All EMNLP 2026 areas and topics

All 16 areas and 100 topics of EMNLP 2026 (5,534 papers in the Main, Findings, Industry and Demo tracks). For each topic you see what the papers do, how many papers it has, its share of all papers in 2025 and 2026, and how many of its papers are in Main and in Findings. The Orals column counts in-person orals.

Areas: Factuality, Evaluation & Expert Domains · LLM Agents · Vision-Language Models · Social Behavior, Bias & Culture · Retrieval & Structured Knowledge · Reasoning & Reinforcement Learning · Efficient LLMs & Decoding · Linguistics, Multilinguality & Translation · LLM Safety & Security · Code, Optimization & Chemistry · Text Generation, Style & Detection · Medical & Clinical NLP · Speech & Audio Models · Fine-Tuning & Preference Alignment · Long-Term Agent Memory · Sequence Model Architectures

Factuality, Evaluation & Expert Domains: 602 papers (10.9% of all, 2025: 12.5%)

These papers study calibration, fact-checking, knowledge editing, benchmarks and annotation, and apply language models to legal, financial, scientific and enterprise documents.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Confidence Calibration and Knowledge Boundaries	These papers study whether models know what they know, covering verbal confidence calibration, belief stability, honesty, overconfidence and hallucination mechanisms.	103	2.32% → 1.86%	57 / 46	13
Scientific Discovery, Writing and Citations	These papers use language models for research ideation, scientific discovery, paper writing and verification, and study how AI changes citation behavior and scientific prose.	63	0.98% → 1.14%	27 / 29	10
Benchmark Construction and Evaluation Reliability	These papers build new evaluation benchmarks across domains and languages and question benchmark validity through contamination detection, compression and diagnostic analysis.	58	1.92% → 1.05%	32 / 24	7
Clarifying Questions and Information-Seeking QA	These papers study how assistants ask clarifying questions, answer information-seeking and misconception-laden queries, and evaluate open QA, mixed with assorted conversation analytics and pragmatics work.	57	1.21% → 1.03%	20 / 25	8
Document Information Extraction and Enterprise QA	These papers extract structured information from corporate, financial and office documents and benchmark question answering over enterprise document collections, with some legal and biomedical extraction mixed in.	52	0.75% → 0.94%	16 / 17	2
Fact-Checking and Factuality Evaluation	These papers verify claims against retrieved evidence, score the factuality of long-form generations and detect fake news, including robustness to conflicting or adversarial evidence.	50	1.12% → 0.90%	26 / 22	6
Human and LLM Data Annotation	These papers study how humans and LLMs annotate data, covering annotator disagreement and label variation, noisy labels, annotation guidelines and cost-aware human-AI labeling pipelines.	47	1.38% → 0.85%	19 / 18	6
Automated Peer Review and Essay Scoring	These papers generate, verify and audit scientific peer reviews with language models, alongside work on multi-trait essay scoring and deep research report evaluation.	43	0.69% → 0.78%	22 / 20	4
Financial NLP Benchmarks and Agents	These papers evaluate and build models and agents for finance, covering financial reasoning benchmarks, retrieval, forecasting, trading, compliance and hallucination detection.	40	0.75% → 0.72%	9 / 20	3
Legal Reasoning and Judgment Prediction	These papers build benchmarks, retrieval systems and reasoning models for legal judgment prediction, legal case retrieval and legal question answering across jurisdictions.	39	0.60% → 0.70%	21 / 14	1
Knowledge Editing and Lifelong Model Editing	These papers update or reverse specific facts and rules inside model parameters, covering batch, multi-hop, temporal and lifelong editing while preserving locality.	29	0.75% → 0.52%	14 / 15	2
Compliance Checking and Auditing Agents	These papers build agentic systems and benchmarks for claim verification, regulatory and advertising compliance checking, fraud detection and evidence-grounded auditing in applied domains.	21	0.06% → 0.38%	2 / 8	3

Orals (65)

• A Contextual Measure of Ambiguity in Prescriptive Texts · PDF
• ABDATABENCH: Evaluating AI Agents for Evidence-Grounded Antibody Record Construction · PDF
• Accelerated Drift of Scientific Prose in the LLM Era: Evidence from an Interrupted Time-Series Analysis · PDF
• AfrIFact: Cultural Information Retrieval, Evidence Extraction And Fact Checking for African Languages · PDF
• AGORA: An Archive-Grounded Benchmark for Agentic Workplace Document Reasoning · PDF
• AnalysisBank: An Expert Analysis Pattern Library for Financial Report Generation · PDF
• Are Benchmarks a Source of Bias in LLMs? MMLU, CommonSenseQA and MGSM under the Microscope · PDF
• BenchBrowser: Retrieving Evidence for Evaluating Benchmark Validity · PDF
• Berry-LR: Literature Retrieval for Automatic Scientific Survey Generation via Berrypicking Model · PDF
• Beyond Dense Retrieval: A Structured-Semantic Approach for Conversation Analytics · PDF
• BiG-SURE - Bipartite Graph for Semantic Uncertainty and Reliability Estimation of LLMs · PDF
• By Their Fruits You Will Know Them: Comparing Formalizations of Law by the Decisions They Encode · PDF
• Can Large Language Models Forecast What Researchers Study Next? · PDF
• Can LLMs Take the Pulse of the Economy? A Real-Time Evaluation of LLM Nowcasts on Macroeconomic Indicators · PDF
• Cited but Not Causal: A Counterfactual Audit of Refusal Explanations in Aligned Language Models · PDF
• ClaimFlow: Tracing the Evolution of Scientific Claims in NLP · PDF
• ConFormBench: Evaluating Evidence-Supported Record Maintenance in Conversational Form Filling · PDF
• Contrastive Attribution in the Wild: An Interpretability Analysis of LLM Failures on Realistic Benchmarks · PDF
• DERELAB: Probing Defeasible Reasoning and Confirmation Bias in LLMs with a Generative Benchmark · PDF
• Deriving Necessary Validity Requirements for Evaluating the Consistency of LLM Responses · PDF
• Detect, Remask, Repair: Diffusion Editing for Faithful Summarization of Evolving Contexts · PDF
• Do Benchmarks Measure the Same Thing? Evaluation-Free Benchmark Redundancy Analysis with SAE · PDF
• Don’t Guess, Ask Wisely: Benchmarking and Training Proactive Elicitation in Open-Domain LLM Conversations · PDF
• Edit Knowledge, Not Just Facts via Multi-Step Reasoning over Background Stories · PDF
• eTracer: Polarity-Aware Evidence Grounding · PDF
• Evaluating Communicative Belief Updates in Large Language Models via Implicature Recognition and Cancellation · PDF
• Evidence Laundering: Deep Research Agents Mistake Correlated Sources for Independent Evidence · PDF
• FinDeepForecast : A Live Multi-Agent System for Benchmarking Deep Research Agents in Financial Forecasting · PDF
• Fixing FOLIO and MALLS: Verified Annotations and an LLM-assisted Framework to Focus Human Relabeling · PDF
• ForensicBench: Evaluating Agentic LLMs on Journal-Entry Fraud Detection · PDF
• Format as Nuisance: Intervention-Inspired Cross-Format Distillation for Knowledge Conflicts · PDF
• Gated Residual Discrepancy for Efficient Bi-Encoder Factuality Verification · PDF
• GLS Chatbot: A Governed Multi-Agent Conversational System for Domain-Specific Customer Service · PDF
• Gold vs. Translation: Are We Measuring Language, Knowledge, or Locality in Translated Benchmarks? · PDF
• Goldsmith: Gold-Loss-Guided Definition Optimization with an Agentic Annotation Harness · PDF
• How broad is that claim? Mapping Generalisation in NLP Research · PDF
• Human Label Variation as Stable Signal: Learning Annotator-Specific Explanation Behavior via Cross-Annotator Preference Optimization · PDF
• KnowledgeGain: Evaluating and Optimizing Science News Generation for Reader Learning · PDF
• LEAP: Likelihood Elicitation and Aggregation for LLM-based Probabilistic Forecasting · PDF
• Learning Diagnostic Clarifying Questions for Conversational Question Answering via Data-Efficient Dynamic Distillation · PDF
• LLM Abstention Can Be a Prompt Artifact, in Addition to Genuine Uncertainty · PDF
• LLM-Based Adversarial Persuasion Attacks on Fact-Checking Systems · PDF
• LLMpedia: A Transparent Framework to Materialize an LLM’s Encyclopedic Knowledge at Scale · PDF
• MAPLE: Metadata Conditioned LLM Pretraining for Locale-Aware Question Answering · PDF
• NARRATOR: Narrative-Aware Routing and Evaluation of Large Language Models · PDF
• Not All Evidence Is Equal: Source-Sensitive Calibration in Large Language Models · PDF
• Popular but Wrong: Understanding and Mitigating LLM Overconfidence through Knowledge Popularity · PDF
• PRISM: Entropy-Guided Multi-View Exploration for Scientific Idea Generation · PDF
• RePair: Turning Retrieval Failures into Counterfactual Hard Pairs · PDF
• Research Design Tracking and Assessment for the Social Sciences · PDF
• Rethinking Clustering as a Conversational Supervision Problem · PDF
• Rethinking Evaluation Practices in Automated Essay Scoring: A Quantitative Analysis over the Past Decade · PDF
• Saying vs. Seeing: Evaluating LLM Conflict Awareness and Resolution Preference under Latent Claim-Evidence Disagreement · PDF
• SciLENS: RL-Driven Autonomous Agents for Scientific Localized Evidence Navigation and Synthesis · PDF
• Show Me How You Cite: Citation Fidelity, Usage, and Topology in the Evidence Behavior of AI Scientific Writing · PDF
• Summarization Is Not Dead Yet · PDF
• TIDE: Tree-Induced Fine-Grained Diagnostic Evaluation from Model Behaviors · PDF
• To Trust or Not to Trust: Retrieval-Augmented Fact Checking in Speech · PDF
• Verify, Then Argue: Claim Verification in Investment Agent Debates · PDF
• What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models · PDF
• What Can Reviews Reliably Tell Us About Books? A Per-Attribute Audit and Tiered Annotated Dataset Across 1.7M Titles · PDF
• What Makes a Good Response? An Empirical Analysis of Quality in Qualitative Interviews · PDF
• Who Annotates in NLP? A Large-scale Assessment of Human Annotation Reporting between 2018 and 2025 · PDF
• Why and When to Use Statistical Significance in Explanation Evaluation: A Systematic Survey and Guidelines for Explainable NLP · PDF
• ZipBench: Low-Cost Framework for Compressing Comprehensive Benchmarks of Large Language Models · PDF

LLM Agents: 594 papers (10.7% of all, 2025: 4.7%)

These papers train and benchmark language agents for tool use, web search, GUI control, long-horizon tasks and multi-agent collaboration, along with skill libraries and agentic reinforcement learning.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Agent Benchmarks and Failure Diagnosis	These papers benchmark and debug LLM agents on real-world planning, enterprise workflows and MCP tool use, and diagnose failures from agent execution trajectories.	98	0.49% → 1.77%	26 / 39	4
Long-Horizon Agents and World Models	These papers train language agents for long-horizon interactive tasks using learned world models, experience reuse, context management and self-evolving policies in environments like ALFWorld and WebShop.	72	0.75% → 1.30%	32 / 40	3
Deep Research and Search Agents	These papers train and evaluate agents that run long-horizon web search for deep research, using reinforcement learning, synthetic search environments and trajectory synthesis.	72	0.26% → 1.30%	30 / 39	3
Agentic RL and Reward Design	These papers improve reinforcement learning for long-horizon agents and reasoning with better credit assignment, rubric and process rewards, rollout budgeting and defenses against reward hacking.	71	0.55% → 1.28%	43 / 23	6
Multi-Agent Collaboration and Communication	These papers design and debug multi-agent LLM systems, covering role and topology adaptation, message pruning, latent communication, failure attribution and robustness to faulty agents.	58	0.77% → 1.05%	34 / 23	9
Tool Use and Function Calling Agents	These papers train and evaluate agents that select, retrieve and call external tools and APIs, with synthetic tool-use data, policy optimization and large tool ecosystems.	55	0.57% → 0.99%	27 / 22	4
Agent Skill Libraries and Skill Evolution	These papers let agents distill, retrieve, revise and evolve reusable procedural skills from execution traces, a topic that is new in 2026.	53	0.00% → 0.96%	21 / 27	4
GUI Agent Benchmarks and Simulators	These papers build benchmarks, simulators and frameworks for agents that operate mobile apps, desktops and web interfaces, testing long-horizon tasks, ambiguity handling and safety.	35	0.17% → 0.63%	18 / 12	4
Proactive Agents and Implicit User Intent	These papers benchmark and train agents that infer implicit user intent, elicit preferences and personalize multi-turn help in shopping, travel planning and assistant settings.	30	0.37% → 0.54%	9 / 17	3
AI Tutoring and Educational Dialogue	These papers build and evaluate adaptive tutoring agents, student simulators and teacher-student dialogue corpora, alongside a smaller group of knowledge distillation papers pulled in by teacher-student wording.	27	0.55% → 0.49%	12 / 9	2
GUI Agent Grounding and RL Training	These papers build benchmarks, GUI grounding methods, world models and reinforcement learning recipes for agents that operate desktop, mobile and browser interfaces.	23	0.26% → 0.42%	12 / 8	0

Orals (42)

• ARM: Role-Conditioned Neuron Transplantation for Training-Free Generalist LLM Agent Merging · PDF
• ATC: Agent Trajectory Control for Role-Contract Enforcement in Black-Box Multi-Agent Systems · PDF
• Beyond Outcome Gaps: Process-Aware Fairness Diagnosis for LLM-based Multi-Agent Decision Systems · PDF
• Causal Abstention for Cost-Aware Language-Agent Collaboration · PDF
• CoPRM: A Collaborative Process Reward Model for Guiding Search Agent Reasoning · PDF
• Decoupling Search from Reasoning: A Vendor-Agnostic Grounding Architecture for LLM Agents · PDF
• Deploying LLM Agents for Auditable Complaint Investigation in Banking · PDF
• DeskCraft: Benchmarking Desktop Agents on Professional Workflows and Human-in-the-Loop Collaboration · PDF
• DigitalCoach: Communication and Grounding Gaps in Human and Agentic Computer Use Coaching · PDF
• DUPO: Dual-Mode Unified Policy Optimization for Tool-Using Agents · PDF
• EDGE: Experience-Distillation for Guided Exploration in Agentic Reinforcement Learning · PDF
• Enrich-Retrieve-Rank: Scaling Capability Discovery Beyond In-Context Routing · PDF
• EvoSkill Injection: Red-Teaming Autonomous Skill Generation and Evolution in Self-Evolving Agents · PDF
• From Task Success to Productive Success: Evaluating Human-AI Collaboration by Quality and Cost · PDF
• HarnessForge: Joint Harness and Policy Evolution for Adaptive Agent Systems · PDF
• HuggingR4: A Progressive Reasoning Framework for Discovering Optimal Model Companions · PDF
• Inferring the Unspoken: Aligning Embodied Agents with Implicit Preferences · PDF
• JarvisGUI: Towards Cross-Device GUI Agents with Dynamic Task Composition · PDF
• Long Live the Librarian! A Persistent Search Sub-Agent for Energy-Efficient Multi-Agent Software Engineering Systems · PDF
• LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis · PDF
• MacAgentBench: Benchmarking AI Agents on Real-World macOS Desktop · PDF
• MARBO: Relational Belief Grounding for LLM Agents in Social Deduction Games · PDF
• MENTOR: Reinforcement Learning via Flexible Teacher-Optimized Rewards for Tool-Use Distillation · PDF
• Mitigating False Credit Propagation: Probabilistic Graphical Reward Aggregation for Rubric-Based Reinforcement Learning · PDF
• Multi-Step Tool-Calling over Korean Open Public APIs: A Benchmark and a Data-Synthesis Recipe · PDF
• OptiMAS: Automatically Optimize Multi-Agent System · PDF
• Reasoning and Tool-use Compete in Agentic RL:From Quantifying Interference to Disentangled Tuning · PDF
• Rethinking the Evaluation of Efficiency Methods for Multi-Agent Systems · PDF
• Robust Multi-Agent LLMs under Byzantine Faults · PDF
• SkillHarm: Lifecycle-Aware Skill-Based Attacks via Automated Construction · PDF
• SkillMe: Evolving Agent Skills for Recurring Personal Workflows · PDF
• SPAR: Experience-Driven Strategy Evolution in Multi-Session B2B Sales · PDF
• Streaming Communication in Multi-Agent Reasoning · PDF
• The Weakest Link Tells It All: Outcome-Supervised Process Reward Modeling via Learnable Credit Assignment · PDF
• TokenCliff: Benchmarking Output-Token Budget Sensitivity in Multi-Turn LLM Agents · PDF
• Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems · PDF
• ToolWeave: Structured Synthesis of Complex Multi-Turn Tool-Calling Dialogues · PDF
• Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills · PDF
• TTML: Test-Time Memory Learning for Label-Free RLVR · PDF
• UserBench: An Interactive Gym Environment for User-Centric Agents · PDF
• When Humans and LLMs Disagree: Understanding Human-LLM Disagreement in Learner Intent Annotation for Educational Chatbot Dialogues · PDF
• World-Model Evolving Agents · PDF

Vision-Language Models: 592 papers (10.7% of all, 2025: 10.8%)

These papers build, benchmark and diagnose vision-language models for images, charts, documents, video, spatial reasoning and embodied action, with work on hallucination and efficient visual tokens.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Vision-Language Model Failures and Robustness	These papers diagnose where vision-language models fail, including language priors overriding perception, framing effects, false premises, miscalibrated confidence and unfaithful visual grounding.	89	1.78% → 1.61%	51 / 38	6
Visually Grounded Multimodal Reasoning with RL	These papers train vision-language models to reason while keeping attention on visual evidence, using reinforcement learning with verifiable rewards, latent visual thinking and visual self-correction.	64	0.60% → 1.16%	35 / 28	5
Video Benchmarks and Text-to-Video Generation	These papers benchmark and improve video-language models on temporal reasoning, captioning, cross-video question answering and text-to-video generation quality.	59	0.83% → 1.07%	32 / 26	6
Omni-Modal Models and Modality Robustness	These papers examine modality bias, uncertainty and robustness in omni-modal and multimodal models, mixed with assorted work on incomplete multimodal learning, emotion recognition and multimodal data pipelines.	57	2.01% → 1.03%	24 / 33	3
Multimodal RAG and Visual Document Retrieval	These papers retrieve images and visually rich documents with multimodal embeddings and late interaction, and use them in multimodal retrieval-augmented generation for long-document question answering.	57	0.80% → 1.03%	34 / 21	8
Vision-Language Hallucination Mitigation	These papers detect and reduce hallucinations, mostly object hallucinations in vision-language models, using attention-head interventions, activation steering, decoding fixes and new benchmarks.	52	1.44% → 0.94%	27 / 25	6
Long and Streaming Video QA	These papers help video language models answer questions about long and streaming videos through frame selection, evidence search and hierarchical multimodal memory.	45	0.43% → 0.81%	24 / 18	2
Vision-Language-Action and Embodied Agents	These papers build and diagnose vision-language-action policies for robot manipulation, navigation and autonomous driving, including grounding, planning and paraphrase robustness.	43	0.60% → 0.78%	21 / 22	4
Visual Token Pruning for Efficient VLMs	These papers cut inference cost in vision-language models by pruning, compressing or condensing visual tokens, with some papers on other efficiency and high-resolution adaptation methods for VLMs.	35	0.66% → 0.63%	17 / 17	3
Spatial Reasoning and 3D Visual Grounding	These papers evaluate and improve the spatial reasoning of vision-language models, including perspective taking, 3D visual grounding, geolocalization and remote sensing.	34	0.52% → 0.61%	15 / 19	3
Chart and Visual Document Reasoning	These papers benchmark and improve vision-language models on charts, diagrams, scientific figures and text-rich images, with a few studies on art, manga and specialized layouts.	33	0.69% → 0.60%	15 / 14	1
Meme Understanding and Harmful Meme Detection	These papers build multimodal models and multilingual benchmarks to interpret memes, humor and cultural references and to detect hateful or harmful memes.	24	0.46% → 0.43%	14 / 10	0

Orals (47)

• Beyond Visual Memory: Mechanistic Diagnostics of Latent Visual Reasoning · PDF
• Can Your Agent Read the Source Faithfully? Benchmarking MLLM Agents on Source-Grounded Multimodal Retrieval · PDF
• ClueTracer: Question-to-Vision Clue Tracing for Training-Free Hallucination Suppression in Multimodal Reasoning · PDF
• Correlation-Flip Visual Injection for Hallucination Mitigation in Large Vision-Language Models · PDF
• Cultural Moment Benchmark: Evaluating Video Cultural Reasoning and Grounding in Southeast Asia · PDF
• Do Vision Language Models Understand Human Engagement in Games? · PDF
• From Content Generation to Learning Support: Pedagogy-Guided Generative Video Tutors for STEM Learning · PDF
• From Global to Regional Coverage: Hyperspherical Prototype Discovery for Storage-Efficient Multimodal Retrieval · PDF
• From Visual Cues to Spoken Narration: Rethinking Audio Description · PDF
• GIFT: Goal-Injected Fine-Tuning for Efficient Manipulation Policy Adaptation · PDF
• HALT: Hallucination Assessment via Log-probs as Time series · PDF
• IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation · PDF
• Intrinsic Temporal Adaptation of CLIP for Partially Relevant Video Retrieval · PDF
• KnowMeBenchV2: Evidence-Grounded Person-Centric Long-Video Understanding · PDF
• Latent Ordinal Evidence, Misaligned Outputs: Inference-Time Ordinal Lens Alignment for Multimodal LLMs · PDF
• LIBERO-Para: A Diagnostic Benchmark and Metrics for Paraphrase Robustness in VLA Models · PDF
• LongNovel: A Multi-Scale Benchmark for Hallucination Detection in Long-Context Novel Summarization · PDF
• MCA: Modality Composition Awareness for Robust Composed Multimodal Retrieval · PDF
• MIDR: Enrichment-Augmented Indexing for Multimodal Document Retrieval · PDF
• MigrationNarrate: A Dataset for Detection of Migration Narratives in YouTube Videos · PDF
• MMEmb-R1: Reasoning-Enhanced Multimodal Embedding with Pair-Aware Selection and Adaptive Control · PDF
• MVP: Multi-View-Guided Token Pruning for Vision-Language Models · PDF
• OmniRetriever: Any-to-Any Audio-Video-Text Retrieval via Fusion-as-Teacher Distillation · PDF
• OVO-S-Bench: A Hierarchical Benchmark for Streaming Spatial Intelligence in Multimodal LLMs · PDF
• Perceive to Hypothesize, Verify to Ground: an Agentic Reasoning Framework for Open-World Geo-Localization · PDF
• PRISM: Self-Pruning Intrinsic Selection Method for Training-Free Multimodal Data Selection · PDF
• PruneGround: Plug-and-play Spatial Pruning for 3D Visual Grounding · PDF
• Rigel: Self-Distilled Score Adaptation for Image and Video Captioning Evaluation · PDF
• ROT: Rotating Hidden States towards Contextual Vectors for Hallucination Mitigation in LVLMs · PDF
• Same Chart, Different Story: Bias in Vision-Language Chart Interpretation · PDF
• Same Trajectory, Contradictory Rewards (RoboRMBench): Paraphrase Fragility in Vision Language Reward Models · PDF
• Seeing with You: Perception-Reasoning Co-evolution for Multimodal Reasoning · PDF
• Shallow to Deep: Aligning Token Pruning with Stage-wise Roles in LVLMs · PDF
• SMART: Sink-based Modality-Aware Redistribution of Transformer Attention · PDF
• State-Conditioned Visual Evidence Retrieval for Fine-Grained Perception in Document Vision-Language Models · PDF
• Think, Look, and Revise: Inconsistency-Aware Visual Self-Correction in MLLMs · PDF
• UEmbed: Unified Sparse and Dense Multimodal Embeddings · PDF
• Understanding the Effects of Distractors on Reasoning Vision-Language Models · PDF
• UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models · PDF
• V-Retrver: Evidence-Driven Agentic Reasoning for Universal Multimodal Retrieval · PDF
• VISA: Video Skeleton-Aware Efficient Frame Selection for Long Video Understanding · PDF
• VisER: Visual Evidence and Reliance for Object Hallucination Detection in LVLMs · PDF
• Vista: An Efficient Late Interaction Engine for Multi-Vector Visual Document Retrieval · PDF
• Weaving Visual Narratives: Agentic Image Bundle Composition Beyond Atomic Visual Matching · PDF
• What Transfers from Text to Vision? Capability Scaling Laws and Transfer Dynamics for VLMs · PDF
• When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs · PDF
• Would you still call this Dax? Novel Visual References in VLMs and Humans · PDF

Social Behavior, Bias & Culture: 550 papers (9.9% of all, 2025: 11.3%)

These papers probe the social, cultural and psychological behavior of language models, covering bias, sycophancy, theory of mind, role-play, personalization, user simulation and LLM judges.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Cognitive Bias, Sycophancy and Moral Judgment	These papers probe language models with methods from psychology and social science, measuring sycophancy, cognitive and social biases, persuasion, moral reasoning and human-model behavioral gaps.	100	2.10% → 1.81%	53 / 46	16
LLM-as-a-Judge Bias and Reliability	These papers test how reliable LLM judges are, exposing position, self-preference and rubric biases and proposing debiasing, calibration and measurement-theory fixes, alongside a few fairness auditing studies.	78	1.03% → 1.41%	39 / 33	8
Role-Playing Agents and Emotional Support Dialogue	These papers train and evaluate role-playing characters, counseling and emotional support agents, and user simulators for multi-turn dialogue, with some general conversational assistant work.	64	1.95% → 1.16%	26 / 34	5
Cultural Knowledge and Cultural Alignment	These papers benchmark and steer how language models understand and adapt to cultural norms, values and region-specific knowledge across many languages and societies.	60	1.21% → 1.08%	25 / 33	5
Social Bias, Stereotypes and Debiasing	These papers measure and trace gender, racial and occupational stereotypes inside language models and debias them, with a share of work on steering vectors and refusal directions more generally.	58	2.01% → 1.05%	37 / 21	5
Personalization and Personality Alignment	These papers adapt model responses to individual users through user profiles, memory selection, preference tracing, persona steering and personality or value alignment.	51	0.77% → 0.92%	20 / 31	4
User Simulators and Human Behavior Simulation	These papers use language models to simulate users, personas and social behavior for evaluating agents and studying surveys and online discourse, and test how faithful these simulations are.	47	0.60% → 0.85%	18 / 18	2
Multi-Agent Debate and Argumentation	These papers improve multi-agent debate for reasoning through better moderation, memory and aggregation, and study persuasion, counterargument generation and competitive debating.	37	0.37% → 0.67%	21 / 14	3
Emotion Recognition and Affective Analysis	These papers recognize, explain and control emotion, sentiment and stance in text and conversation, and probe how models represent affect internally.	29	0.80% → 0.52%	22 / 7	4
Theory of Mind and Social Intelligence	These papers benchmark theory-of-mind and social reasoning in text and multimodal models, from multi-party meetings and body language to deception and social deduction games.	26	0.46% → 0.47%	13 / 13	3

Orals (55)

• AgentJudgeBench: A Multi-Difficulty Benchmark for Evaluating LLM Judges on Agentic Tool-Calling · PDF
• Aligned but Not Partner-Specific: How Multimodal LLM Agents Succeed in Reference Games Without Forming Conceptual Pacts · PDF
• Auditing Intersectional Bias in Clinical LLMs: Sparse Autoencoders Reveal Race-HIV Stigma Circuits · PDF
• BenchPreS: A Benchmark for Context-Aware Personalized Preference Selectivity of Persistent-Memory LLMs · PDF
• Causal Interventions on Continuous Variables: A Case Study on Verb Bias in Steering Vectors for In-Context Learning · PDF
• Creative Generation via Multi-Agent Debate: Does Debate Suppress Diversity? · PDF
• Cross-Centaur: Modeling Individual Decision-Making Across Psychological Tasks with Language Models · PDF
• Do Large Language Models Adapt Politeness Like Humans? · PDF
• Does Sensitivity to Pragmatic Norms Emerge with More Parameters, More Data, or More Conversations? · PDF
• DuplexGen: Adaptive Synthesis of Human-AI Turn-Taking Dialogues · PDF
• EmoStance: Response-Side Affective-Orientation Control for Empathetic Response Generation via Emoji Weak Supervision · PDF
• Forewarned Is Not Forearmed: How Coordinated LLMs Shift Reader Opinions in News Comment Sections · PDF
• From Chatbots to Confidants: A Cross-Cultural Study of LLM Adoption for Emotional Support · PDF
• From Debate to Decision: Conformal Social Choice for Safe Multi-Agent Deliberation · PDF
• From Monolithic Blending to Agentic Orchestration: Dynamic Response for Conversational Assistants at Scale · PDF
• How Identity and Opinion Shape Political Sycophancy in LLMs · PDF
• How to Solve Noise and Hard Samples: Filtering and Difficulty-adaptive Augmentation for Fine-grained Negative Emotion Classification · PDF
• HugAgent: A Human Simulation Benchmark for Individual-Level Reasoning · PDF
• Identifying and Mitigating Bottlenecks in Role-Playing Agents: A Systematic Study of Disentangling Character Profile Axes · PDF
• Justification of Emotional Appeal in Computational Argumentation · PDF
• Less Is Moral: A CHARMing Framework for Moral Foundations Detection in Endorsement Behaviour · PDF
• Lessons Without Borders? Evaluating Cultural Alignment of LLMs Using Multilingual Story Moral Generation · PDF
• LLUMI: Improving LLM Writing Assistance for Mental Health Support with Online Community Feedback · PDF
• Lost in Delusion: Examining LLM Safety Under User Delusions and Distress · PDF
• Measuring the Behavioral Fidelity of Long-Horizon Human Activity Simulations · PDF
• MeetingToM: Evaluating Multimodal LLMs on Theory-of-Mind Reasoning in Multi-Party Meetings · PDF
• MMDS-Bench: Benchmarking Multimodal Large Language Models on Dynamic Stance in Social Media Interactions · PDF
• Moral Lenses, Political Coordinates: Towards Ideological Positioning of Morally Conditioned LLMs · PDF
• Multi-Expert Conformal Risk Control for Pairwise LLM Judging in Open-Ended Dialogue · PDF
• Open Human Value Mining: Constrained Argument Extraction and Value Concept Induction with LLMs · PDF
• Persona-CAPS: Controllable and Auditable LLM Personality via Retrieval over Situation–Behavior Contingencies · PDF
• Physiognomy Revisited: How Vision–Language Models Encode Sensitive Attributes Through Image Transformation · PDF
• PopResume: Causal Fairness Evaluation of LLM/VLM Resume Screeners with Population-Representative Dataset · PDF
• Predicting Long-Term Conversational Skill Development in Mental-Health Crisis Counselors · PDF
• Race in the Record: Measuring Racial Signal and Model Inference in Police Arrest Narratives · PDF
• Rating the Pitch, Not the Product: User Evaluations of LLMs Reflect Expectations More Than Performance · PDF
• RAW: Refusal Aversion in the Wild, A Causal Measurement Method for Deployed LLMs · PDF
• Reward-Decomposed Reinforcement Learning for Immersive Video Role-Playing · PDF
• Security and Privacy Prompts in the Wild: What Users Ask LLMs and How LLMs Respond · PDF
• SIC-Agents: Benchmarking and Building Adaptive Simulator for Pediatric Serious Illness Communication Training · PDF
• SocialArena: Evaluating Social Intelligence in Large Language Models · PDF
• SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generation · PDF
• Surfacing the Unsaid: CUE-Bench for Affective Stance in Chinese Discourse · PDF
• Tampered ML: Evaluating LLM Judgment of Machine Learning Research · PDF
• Text as a Universal Interface for Transferable Personalization · PDF
• TukaBench: A Culturally Grounded Jailbreak Benchmark for African Languages · PDF
• What Are We Measuring in NLG? A Meta-Analysis of Evaluation Trends 2020-2025 · PDF
• What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal · PDF
• When Calibration Rankings Reverse: Accuracy-Controlled Evaluation for Fair Comparison of LLMs · PDF
• When LLMs Benchmark Themselves: Deconstructing Self-Bias in Automated Evaluation · PDF
• When to Trust: Margin-Calibrated Routing Between LLM-as-a-Judge and Human Experts · PDF
• Whose Norms? Disentangling Cultural and Personal Alignment in Large Language Models · PDF
• XCR-Bench: Benchmarking Cross-Cultural Reasoning in LLMs via Culture-Specific Items and Hall’s Triad · PDF
• You Really Didn’t Get That? Benchmarking Social Pragmatic Inference for Indirect and Playful Chinese Online Comments · PDF
• You Shouldn’t Have Asked: A Pragmatics-Inspired Taxonomy for Evaluating LLM Refusals · PDF

Retrieval & Structured Knowledge: 483 papers (8.7% of all, 2025: 11.2%)

These papers improve retrieval and retrieval-augmented generation and reason over knowledge graphs, tables, databases, entities and events, including recommendation systems.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Retrieval-Augmented Generation and Evidence Grounding	These papers improve how RAG systems select, trust and cite retrieved evidence, covering knowledge conflicts, poisoning robustness, attribution and RAG evaluation.	102	2.41% → 1.84%	44 / 45	3
Dense Retrieval and LLM Reranking	These papers train dense and generative retrievers and LLM rerankers, with a focus on reasoning-intensive and efficient text retrieval.	77	1.41% → 1.39%	36 / 27	6
Knowledge Graph Reasoning and Link Prediction	These papers reason over knowledge graphs and text-attributed graphs, covering KGQA, temporal knowledge graph forecasting, link prediction and graph serialization for language models.	60	1.61% → 1.08%	32 / 26	1
Graph RAG and Multi-Hop QA	These papers answer multi-hop questions by retrieving evidence over knowledge graphs, hypergraphs and document graphs, often with agentic or iterative retrieval.	54	1.00% → 0.98%	22 / 29	1
Entity Linking, Relation Extraction and Taxonomies	These papers link entities to knowledge bases, extract relations and named entities, and induce taxonomies, mixed with assorted work on topic modeling, embedding geometry and intent discovery.	48	1.72% → 0.87%	22 / 23	9
Generative and Sequential Recommendation	These papers build LLM-based recommender systems that generate semantic item IDs, model user histories and preferences, and suggest queries in e-commerce search.	45	0.83% → 0.81%	21 / 13	2
Text-to-SQL and Database Querying	These papers translate natural language into SQL with schema linking, chain-of-thought training, reinforcement learning and candidate selection, and build more realistic text-to-SQL benchmarks.	39	0.75% → 0.70%	16 / 17	1
Table Reasoning and Table Question Answering	These papers answer questions and reason over tables, including multi-table retrieval, SQL-guided decomposition, table graphs and multimodal table images.	36	0.75% → 0.65%	24 / 10	3
Event Extraction and Time-Series Grounding	These papers extract events, arguments and timelines from documents and connect language with time-series data for event detection, forecasting and temporal explanation.	22	0.75% → 0.40%	7 / 14	0

Orals (26)

• Auditing RAG Necessity: A Conformal Test for Source-Independent Answers · PDF
• CROG: Coverage-aware Reasoning Optimization on Graphs for Multi-hop Question Answering · PDF
• Cross-lingual Biography Enrichment via Claim Extraction and Alignment · PDF
• DiffIE: Diffusion-based Open Information Extraction · PDF
• From Prompts to Trees: Effective LLM-Guided Tree Generation for Few-Shot Tabular Classification · PDF
• INTRYGUE: Induction-Aware Entropy Gating for Reliable RAG Uncertainty Estimation · PDF
• Iterate Until Retrieved: Factual Nugget Optimization for Discoverable Continual Corrections in Agentic RAG · PDF
• LaPR: Latent Preference Reasoning for Generative Personalized Retrieval · PDF
• Latent Bridges for Multi-Table Question Answering · PDF
• LTEM: LLM-Friendly Table-Enhanced Markdown for Table Understanding and Generation · PDF
• Measuring Data Diversity with Embeddings: A Taxonomy and Benchmark for Diversity Measures · PDF
• MEMORA: A Memory-Enhanced Multimodal Committee Reranking Agent for Reasoning-Intensive Retrieval · PDF
• MICE: Minimal Interaction Cross-Encoders for efficient Re-ranking · PDF
• NE-R1: Enhancing Named Entity Recognition Model via Reinforcement Learning · PDF
• Ontology-Based Knowledge Graphs for Industrial Standards with Scope-Aware Logical Rule Structuring · PDF
• Ontology-Driven Structural Regularization for Document-Level Relation Extraction · PDF
• RCEM: Robust Conversational Search EMbedder in Distributional Shift · PDF
• SafeGEO: Understanding Generative Engine Optimization Risks in Recommendation Agents · PDF
• SHIFT: Self-reconstruction Harnesses Implicit Fine-grained Thinking for Retrieval · PDF
• SPARROW: Scalable Taxonomy Induction via Structure-Preserving Partitioning and Constraint-Guided Merging · PDF
• TEMPO: Realistic Multi-Domain Benchmark for Temporal Reasoning-Intensive Retrieval · PDF
• TETHER: Training-Free, Elastic, ToC-Anchored Hierarchy Extraction and Recovery in Design-Driven Long Documents · PDF
• Towards Effective Structured Context Modeling for Conversational Recommender Systems via Dual-node Monte Carlo Tree Search · PDF
• Typed Relation Path Geometry in Sentence Embedding Spaces · PDF
• Variable-size Discovery and Representation of Atomic Propositions · PDF
• When the Bottleneck Shifts: Diagnosing and Closing Information Gaps in Interactive Text-to-SQL · PDF

Reasoning & Reinforcement Learning: 452 papers (8.2% of all, 2025: 8.0%)

These papers improve and analyze model reasoning through reinforcement learning with verifiable rewards, efficient chain-of-thought, mathematical problem solving and faithfulness diagnostics.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Efficient and Latent Chain-of-Thought Reasoning	These papers cut overthinking in reasoning models through chain-of-thought compression, adaptive token budgets, early stopping, parallel thinking and latent or implicit reasoning.	139	2.47% → 2.51%	62 / 75	9
Chain-of-Thought Faithfulness and Reasoning Diagnostics	These papers analyze why chain-of-thought reasoning fails, whether reasoning traces are faithful, and how to benchmark logical and causal reasoning, alongside work on overthinking and test-time compute.	103	1.06% → 1.86%	48 / 52	9
RLVR and On-Policy Distillation	These papers refine reinforcement learning with verifiable rewards and on-policy distillation for reasoning, tackling credit assignment, entropy collapse, advantage estimation and label-free rewards.	89	0.09% → 1.61%	39 / 50	8
Mathematical Reasoning and Formal Theorem Proving	These papers improve and evaluate mathematical problem solving through verifiers, process reward models, distractor robustness, autoformalization and formal proof benchmarks.	68	1.87% → 1.23%	27 / 40	4
Self-Improving Reasoning and Self-Correction	These papers improve model reasoning through self-evolution, reinforcement learning with guided exploration, latent reasoning and activation steering, alongside some work on calibration and abstention.	53	2.53% → 0.96%	33 / 20	7

Orals (37)

• Bayesian Sparse Low-Rank Adaptation for Large Language Model Uncertainty Estimation · PDF
• Beyond Append-Only Decoding: A Recursive Edit-Based Encoder for Arithmetic Length Generalization · PDF
• Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies · PDF
• CoDiQ: Test-Time Scaling for Controllable Difficult Question Generation · PDF
• Cognitively Decoupled Meta-Induction for Out-of-Distribution Generalization · PDF
• Conformal Certification of Reasoning Trace Prefixes · PDF
• ConsensusBench: Benchmark of Consensus Nodes for LLM Reasoning via Outcome Reward Densifying · PDF
• Correct Prediction, Wrong Steps? Consensus Reasoning Knowledge Graph for Robust Chain-of-Thought Synthesis · PDF
• Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following · PDF
• Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding · PDF
• Don’t Peek at the Answer: Outcome-Masked Group Relative Policy Optimization for Label-Free RLVR · PDF
• Escaping the KL Agreement Trap in On-Policy Distillation · PDF
• From Rollouts to Recipes: Self-Contained Post-Training for LLMs · PDF
• Gradients Know What Outcomes Don’t: Unlocking Reinforcement Learning for LLM Reasoning with Gradient-Aligned Rewards · PDF
• Improving Latent Reasoning in LLMs via Soft Concept Mixing · PDF
• In-Place Feedback: Reliable Refinement for Multi-Turn Expert-LLM Collaboration · PDF
• Last Layer Logits to Logic: Empowering LLMs with Logic-Consistent Structured Knowledge Reasoning · PDF
• LatentRevise: Learning from Zero-Hit Reasoning · PDF
• Learning How to Align: Elastic Domain Adaptation for Improving Reasoning Robustness of Large Language Models · PDF
• LogicGraph: Benchmarking Multi-Path Logical Reasoning via Neuro-Symbolic Generation and Verification · PDF
• MathDebugger: A Benchmark for Type-Aware Error Detection in Synthetic Math Data · PDF
• Multi-View Structural Interpretability of Agentic Reasoning Traces · PDF
• Policy of Thoughts: Scaling LLM Reasoning via Test-time Policy Evolution · PDF
• ReasoningFlow: Discourse Structures for Understanding LLM Reasoning Traces · PDF
• Reinforcement Learning Improves Traversal of Parametric Knowledge in LLMs · PDF
• Robust Failure, Conservative Repair: Textual Knowledge Distillation from Cross-Model Failures · PDF
• Sample Where You Struggle: Sharpening Base Model Reasoning via Entropy-Guided Power Sampling · PDF
• Share More, Search Less: Collaborative Parallel Thinking for Efficient Test-Time Scaling · PDF
• SSR: Symbolic-Neural Soft Reasoning for Robust and Verifiable Thinking Chains · PDF
• The Imitation Game: When LLMs Learn to Reason Like Programs via Code-Centric Reasoning Data Synthesis · PDF
• The Importance of Being Statistically Earnest: A Critical Re-evaluation of GSM-Symbolic · PDF
• TopoAlign: A Framework for Aligning Code to Math via Topological Decomposition · PDF
• Towards Understanding Pause Token Fine-Tuning Dynamics: A Mode Retention Perspective · PDF
• Training-Free Chunk-Level Guided Generation for Mathematical Reasoning · PDF
• When Must Faithful Chain-of-Thought Fail? An Information-Theoretic Bound and Its Empirical Limits · PDF
• When Not to Think: Mitigating the Overthinking Effect in Content Moderation · PDF
• When VLMs ’Fix’ Students: Identifying and Penalizing Over-Correction in the Evaluation of Multi-line Handwritten Math OCR · PDF

Efficient LLMs & Decoding: 422 papers (7.6% of all, 2025: 6.8%)

These papers make language models faster and smaller through pruning, quantization, KV cache compression, mixture-of-experts, speculative decoding and diffusion language models, along with work on unlearning.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Model Pruning and Sparsity	These papers compress language models through structured pruning, quantization and distillation and measure real inference speedups, while the cluster also mixes in assorted training, data selection and representation studies.	127	3.47% → 2.29%	66 / 48	11
Diffusion Language Models and Parallel Decoding	These papers build, scale and speed up masked and continuous diffusion language models, with new remasking, parallel decoding, length control and distillation strategies.	72	0.34% → 1.30%	38 / 33	6
KV Cache Compression and Eviction	These papers shrink the key-value cache for long-context inference through token eviction, quantization, low-rank approximation and adaptive per-layer budgets.	57	0.52% → 1.03%	41 / 14	6
Speculative Decoding and Inference Acceleration	These papers speed up language model inference with draft models, draft trees, retrieval-based drafting and smarter verification, plus related work on efficient sampling.	42	0.69% → 0.76%	25 / 17	5
Low-Bit Weight and Activation Quantization	These papers compress model weights, activations and KV caches with post-training, mixed-precision, ternary and quantization-aware methods that handle outliers.	37	0.52% → 0.67%	21 / 13	1
Machine Unlearning in Language Models	These papers remove targeted knowledge from text and multimodal models while preserving utility, and test how robust the forgetting is to relearning and recovery attacks.	31	0.63% → 0.56%	12 / 18	1
MoE Compression and Efficient Training	These papers compress, merge and quantize mixture-of-experts models and make MoE training and fine-tuning more efficient, alongside some work on the Muon optimizer and parallelism.	29	0.26% → 0.52%	17 / 11	3
Mixture-of-Experts Routing and Fine-Tuning	These papers improve how MoE models route tokens to experts, prune or reduce experts at test time, and fine-tune MoE models without breaking their routing structure.	27	0.40% → 0.49%	14 / 13	1

Orals (34)

• A Comparative Analysis of LLM Memorization at Statistical and Internal Levels: Cross-Model Commonalities and Model-Specific Signatures · PDF
• Accelerating Constrained Decoding with Token Space Compression · PDF
• Beyond Magnitude: Contrastive Routing for Modular Mixture-of-Experts · PDF
• Breaking the Likelihood Trap: Variance-Calibrated Modulation for Large Language Model Decoding · PDF
• CacheSpec: Finding the Sweet Spot for Small Models in Large Language Models · PDF
• CAVEWOMAN: How Large Language Models Behave Under Linguistic Input and Output Compression · PDF
• Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding · PDF
• Evo: Autoregressive–Diffusion Large Language Models with Evolving Balance · PDF
• Fast Block Diffusion Training via Multi-Mask Context Sharing · PDF
• Follow the Latent Roadmap: Navigating Revocable Decoding for Diffusion LLMs with Anchor Tokens · PDF
• Forget Without Compromise: Nexus Sampling for Streaming KV-Cache Eviction Under Fixed Budgets · PDF
• Generation-Gated Cascaded Routing for Large Language Models · PDF
• Graph-Guided Selective Unlearning for Language Models: Controlling Support Routes Beyond Forget Seeds · PDF
• How Should LLMs Consume High-Quality Data? Optimal Data Scheduling via Quality-Aware Functional Scaling Laws · PDF
• HTAM: Hierarchical Transition-Attended Memory for Operator Optimization · PDF
• Last But Not Least: Boundary Attention CalibratiON for Multimodal KV Cache Compression · PDF
• Manifold Embedding: A Point-to-Hyperplane Approach · PDF
• Micro Language Models Enable Instant Responses · PDF
• Multi-Block Diffusion Language Models · PDF
• Muon2: Boosting Muon via Adaptive Second-Moment Preconditioning · PDF
• No Training, No Architecture Change: Per-Cell Warm-Up Calibration for LLM Agent Memory · PDF
• Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models · PDF
• PRISM: Prompt Representation Internalization via Selective Compression into LoRA Weights · PDF
• Pruning Laws for Large Language Models · PDF
• ResCode: Accurate Expert Compression of Sparse MoE Models via Residual-code Restoration · PDF
• RoleKV - Role-Aware KV Cache Management for the Inverted Age-Importance of LLM Agent Context · PDF
• Scaling Laws for Mixed Quantization · PDF
• SlimKV: Joint Token-Feature KV Cache Compression with Reconstruction-Free Beacon Attention · PDF
• Spexis: Speculative Lookahead Scheduling for LLM Inference · PDF
• Step-Wise Refusal Dynamics in Autoregressive and Diffusion Language Models · PDF
• UT-ACA: Uncertainty-Triggered Adaptive Context Allocation for Long-Context Inference · PDF
• VIB-Simp: IB-Inspired Output-Logit Regularization for German Text Simplification · PDF
• What Matters for Aggressive Decoding-Time KV Eviction? Temporal Aggregation and Ranking Preservation · PDF
• WnW: Waxing-and-Waning KV Cache for Long-Form Speech LLMs · PDF

Linguistics, Multilinguality & Translation: 390 papers (7.0% of all, 2025: 9.8%)

These papers compare models with human language processing, study cross-lingual transfer, dialects and tokenization, and improve and evaluate machine translation.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Psycholinguistics and Linguistic Knowledge Probing	These papers compare language models with human sentence processing and test their syntactic, lexical and typological knowledge, with a mix of morphology, parsing and corpus work.	120	2.24% → 2.17%	70 / 48	25
Cross-Lingual Transfer and Multilingual Representations	These papers measure and close the gap between English and other languages through cross-lingual alignment, representation analysis, multilingual judges and multilingual training data.	101	2.99% → 1.83%	54 / 45	12
Dialects, Scripts and Underrepresented Languages	These papers study dialects, writing systems and romanized or cross-script text in Arabic, Indic, African and other low-resource languages, alongside a mix of other language-specific resources and benchmarks.	64	1.92% → 1.16%	34 / 27	11
Document-Level and Low-Resource Translation	These papers study how language models translate, build translation benchmarks and metrics, and improve document-level and low-resource machine translation with reasoning and reinforcement learning.	61	1.23% → 1.10%	32 / 29	8
Machine Translation Quality Evaluation	These papers evaluate machine translation with LLM judges, error span detection, document context and human studies, including literary and culturally loaded translation.	24	0.98% → 0.43%	8 / 14	2
Tokenization and Subword Vocabularies	These papers analyze and redesign tokenizers, including byte-level, grapheme and morphology-aware subwords, and study their effect on multilingual and low-resource modeling.	20	0.46% → 0.36%	9 / 11	5

Orals (63)

• A Scaled-Up Empirical Study of Syntactic Language Models · PDF
• AdiDhwani: Preservation-Oriented ASR for Endangered Bhili and Mundari Languages · PDF
• AfriSUD: A Dependency Treebank Collection for Evaluating Models on African Languages · PDF
• An Interpretable Linguistically-Grounded Analysis of Scene Saliency in Movie Screenplays · PDF
• Arabic Morphosyntactic Tagging and Dependency Parsing with Large Language Models · PDF
• Beetle: A Bilingual Model Suite for Modelling Second-Language Processing · PDF
• Beyond Accuracy: Community Perspectives on Machine Translation · PDF
• Beyond “To whom it may concern”: Tailoring Machine Translation to Audience and Intent · PDF
• Can Large Language Models Handle Discourse Particles? A Case Study of Colloquial Malay · PDF
• Chronoberg: A 250-Year Diachronic Corpus with Temporal Affective Lexicons for Language Model Analysis · PDF
• Cross-Lingual Representation Alignment by Token-Level Optimal Transport in a Language-Agnostic Space · PDF
• Cross-lingual Representation Learning via Centroid Intervention Fusion · PDF
• CWoMP: Morpheme Representation Learning for Interlinear Glossing · PDF
• Disentangling Spurious Correlations: How IRM Handles Diverse Shortcuts in Text Classification · PDF
• Disentangling Statistical Preemption from Entrenchment in Language Models’ Avoidance of Overgeneralization · PDF
• Do Language Models Understand Implicit Logical Meaning? A Case Study of Scope Ambiguity Resolution in Context · PDF
• Generative Models Enhanced by Sequence Labelling and Aspect-Code Switching Improve Cross-lingual Aspect-Based Sentiment Analysis · PDF
• HalloMTBench: A Large-Scale Human-Verified Benchmark for Diagnosing Hallucination Triggers in LLM Translation · PDF
• HOMURA: Taming the Sand-Glass for Time-Constrained LLM Translation via Reinforcement Learning · PDF
• How Do Language Models Represent and Use Phonological Information for Allomorph Selection? · PDF
• How Human-Like Are Large Language Models? A Register-Aware Linguistic Evaluation Framework · PDF
• IdioLink: Retrieving Meaning Beyond Words Across Idiomatic and Literal Expressions · PDF
• Idioms Understood, Yet Translated Literally: Diagnosing Literal Translation Bias in Multilingual LLMs · PDF
• INCLUDE-2.0: Disentangling Knowledge Gaps in Multilingual LLM Evaluation · PDF
• IndoDialect: A Multi-Dialect Conversational Dataset for Indonesian Local Languages · PDF
• LangMAP: A Language-Adaptive Approach to Tokenization · PDF
• Language Identification for Non-Standard Varieties: Language Varieties as Geolects on a Geodesic Grid · PDF
• Language Proficiency Assessment from Eye Movements in Naturalistic Passage Reading · PDF
• Left-Branching Transformers Excel at Right-Branching Languages: Data Shapes Word Order Preferences in Language Models · PDF
• Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection · PDF
• Lossy Orthography and Linguistic Inference: Computational Approaches to Arabic Rasm Restoration · PDF
• Not all Negation Cues are Equal: Affixal Negations Yield Better Negation Understanding · PDF
• Objective vs. Search: Decomposing What Makes a Good Tokeniser · PDF
• Omnilingual MT: Machine Translation for 1,600 Languages · PDF
• On the Optimality of Kinship Naming: an Information-theoretic Approach · PDF
• PAMT: Process-Aligned Reinforcement Learning for Multi-Domain Machine Translation · PDF
• Parallel Tokenizers: Rethinking Vocabulary Design in Cross-Lingual Transfer of Low-Resource Languages · PDF
• PolERo: Studying Political Evasion in Romanian · PDF
• PunGraph: Retrieval-Enhanced Phonetic-Semantic Graph Reasoning for Pun Understanding · PDF
• Rashid: A Cipher-Based Framework for Exploring In-Context Language Learning · PDF
• Repetition over Diversity: High-Signal Data Filtering for Sample-Efficient German Language Modeling · PDF
• Revisiting the Estimation of Unigram Surprisal in Computational Psycholinguistics · PDF
• Script Is Not Preprocessing: Tokenizer-Mediated Transfer in Endangered Language Machine Translation · PDF
• Self-Evolving Optimization of Agentic Systems for Machine Translation · PDF
• SemBridge: Language Transfer in Sparse Encoders via Multilingual Semantic Bridge · PDF
• Size Sound Symbolism Survives Genealogy Across 27 Languages · PDF
• Syntactic Belief Update as the Driver of Garden Path Processing Difficulty · PDF
• The Emergence and Sparsification of a Syntactic Circuit · PDF
• The Geometry of Low-Resource Language Representations · PDF
• The Interlingua Hypothesis: LLMs Translate via a Latent Task-agnostic Feature Space · PDF
• The Multilingual FrameNet Corpus · PDF
• Tokenization as Structural Inductive Bias: Grapheme-Constrained Subwords for Multilingual Modeling and Vocabulary Surgery · PDF
• Tracing Multilingual Representations in LLMs with Cross-Layer Transcoders · PDF
• Tracing the complexity profiles of different linguistic phenomena through the intrinsic dimension of LLM representations · PDF
• Transporting Concepts Across Languages: Optimal Transport Toward English-Level Performance in Multilingual LLMs · PDF
• Using LMs to Model the Effects of Context and Coreference during Sentence Comprehension · PDF
• Using Prosody to Predict Syntactic Structure · PDF
• VAA-CSEC: Vote-guided Advantage Allocation for Chinese Semantic Error Correction · PDF
• What Transfers Across Cuneiform? Script-Level Substrate and Language-Specific Regularities via Segmentation and Representation · PDF
• When Does Auxiliary Classification Help Regression in Practice? A Case Study in Arabic Dialectness Modeling · PDF
• When Similar Means Different: Evaluating LLMs on Arabic–Hebrew Cognates · PDF
• When Tokenization is Secretly Output Supervision · PDF
• Who Counts as "the People"? Measuring Contested Political Concepts from Political Speech using Masked Language Models · PDF

LLM Safety & Security: 386 papers (7.0% of all, 2025: 7.1%)

These papers attack and defend language models and agents through jailbreaks, prompt injection, backdoors, watermarking, privacy protection and safety alignment.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Safety Alignment and Over-Refusal	These papers study how safety alignment and refusal work inside models, defend against jailbreaks and over-refusal, preserve safety during fine-tuning and benchmark content moderation.	93	1.15% → 1.68%	57 / 32	12
Jailbreak and Backdoor Attacks	These papers design jailbreak and backdoor attacks on language and vision-language models and defend against them with detection probes, guardrails and alignment interventions.	77	1.58% → 1.39%	35 / 42	4
Agent Safety and Tool-Use Risks	These papers benchmark and guard tool-using agents against unsafe actions, prompt injection, malicious skills, privacy leaks and backdoors during multi-step execution.	50	0.37% → 0.90%	24 / 20	0
Indirect Prompt Injection Attacks	These papers attack and defend tool-using agents against indirect prompt injection and build red-teaming methods and benchmarks for agent security.	39	0.72% → 0.70%	13 / 24	2
Privacy Leakage, PII and Membership Inference	These papers measure and mitigate privacy risks through PII detection, text sanitization, re-identification audits, membership inference and contextual integrity in agents.	37	0.66% → 0.67%	17 / 14	5
Detoxification and Moderation Evasion	These papers reduce toxic generations, generate counterspeech and study how harmful content, fake news and poisoned corpora evade detectors, with a few loosely related robustness papers mixed in.	35	1.66% → 0.63%	22 / 13	4
Multimodal and Text-to-Image Safety	These papers benchmark and defend multimodal and text-to-image models against unsafe generation, jailbreaks and culturally specific risks, using safety unlearning, detoxification and guard models.	31	0.46% → 0.56%	13 / 15	0
Text Watermarking and Model Fingerprinting	These papers embed and detect watermarks in generated text and fingerprint models for copyright protection, balancing detectability, robustness to edits and text quality.	24	0.55% → 0.43%	12 / 12	1

Orals (28)

• A Unified Mechanistic Analysis of Knowledge- and Safety-Based Refusals · PDF
• Beyond Semantic Accuracy: Consequence-Aware Evaluation for Safety-Critical Language Understanding · PDF
• CASTLE: A Comprehensive Benchmark for Evaluating Student-Tailored Personalized Safety in Large Language Models · PDF
• CopyShield: A Cross-Level Benchmark of Copyright Defenses in LLMs · PDF
• GraphProfiler: Source-Linked Sensitive Attribute Inference via Personal Knowledge Graphs · PDF
• Harm Laundering in GPT Models: Evidence That Gender Discrimination Is Transformed Rather Than Reduced Across Safety-Trained Generations · PDF
• "I Strongly Suspect This Website Is a Scam": Benchmarking PII Leakage and Detection without Defense in Autonomous Web Agents · PDF
• InterPol: De-anonymizing LM Arena via Interpolated Preference Learning · PDF
• JADES: A Universal Framework for Jailbreak Assessment via Decompositional Scoring · PDF
• Legible but Not Localized: Jailbreak Geometry in Mixture-of-Experts Language Models Is Attack-Specific · PDF
• Mitigating LLM Over-Refusal via Dynamic Semantic Routing Calibration · PDF
• Mood Matters: How Syntactic Sensitivity Undermines Safety Alignment · PDF
• Multilingual Safety Signals Are Multi-Layered: Filtering Safety-Degrading Data for Safer LLMs · PDF
• On the Sensitivity of Instruction-tuned LLMs to Harmful Sentences in Long Inputs · PDF
• One Refusal Isn’t Enough: Multi-Turn Jailbreak Attacks on Large Language Models · PDF
• Only Pay What You Must Spend: On-Demand Privacy Budget Payment for Differentially Private RAG · PDF
• Q-Align: Query-Boundary Multilingual Alignment for LLM Safety Guardrails · PDF
• SafeSteer: Localized On-Policy Distillation for Efficient Safety Alignment · PDF
• Safety Paradox: How Enhanced Safety Awareness Leaves LLMs Vulnerable to Posterior Attack · PDF
• Self-Sanitize: Cognitive-Inspired In-Place Mitigation of LLM Harmful Responses · PDF
• T-MAP: Red-Teaming LLM Agents with Trajectory-aware Evolutionary Search · PDF
• The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs · PDF
• The Persona-Switch Jailbreak Backdoor in LLMs: From Payload Retrieval to Trigger-Gated Persona-State Selection · PDF
• Thunder-DeID+: Boundary-Aware Korean De-identification from Court Judgments to Broader Sensitive Documents · PDF
• TRACE: An Evidence-Grounded Benchmark for Safety Evaluation of Large Reasoning Models · PDF
• VerTox: Verifiable Reward-Guided Corpus Poisoning Against Neural Ranking Models · PDF
• When Hate Meets Facts: LLMs-in-the-Loop for Check-worthiness Detection in Hate Speech · PDF
• Will the User Ever Know? Covert Indirect Prompt Injection Attacks on Tool-Using LLM Agents · PDF

Code, Optimization & Chemistry: 199 papers (3.6% of all, 2025: 2.5%)

These papers use language models and agents to write, test and repair code, solve software engineering tasks, discover algorithms through evolutionary search and design molecules and drugs.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Code Generation, Testing and Repair	These papers train and evaluate code models with unit-test generation, execution feedback, reward models and debugging to make generated code functionally correct.	76	1.26% → 1.37%	29 / 44	5
Coding Agents for Software Engineering	These papers benchmark and train coding agents on repository-level tasks such as issue localization, program repair, code retrieval and refactoring.	48	0.26% → 0.87%	24 / 20	7
Chemistry, Molecules and Drug Discovery	These papers apply language models and agents to molecular property prediction, molecular optimization, retrosynthesis, structure elucidation, protein design and drug discovery.	40	0.80% → 0.72%	13 / 25	2
Optimization Modeling and Evolutionary Algorithm Discovery	These papers use language models to formulate operations research and optimization problems and to evolve algorithms, code and designs through search-based self-improvement loops.	35	0.17% → 0.63%	17 / 18	3

Orals (17)

• Bayesian Control for Coding Agents · PDF
• Beyond the Vacuum: Combinatorial Strategy Selection for Competitor-Aware Generative Engine Optimization · PDF
• Code-TVT: Synthesizing Test-Case-Driven Verifiable Tasks for Code Reinforcement Learning · PDF
• CodeTD: Topology of Attention Detects Hallucinations in Code LLMs · PDF
• Compositional Machine Design as Program Synthesis with LLMs · PDF
• CORE-Bench: A Comprehensive Benchmark for Code Retrieval in the Era of Agentic Coding · PDF
• Dockerless: Environment-Free Program Verifier for Coding Agents · PDF
• Entropy-Regularized Rank-Masked Policy Optimization for Test-Time Reinforcement Learning in Code Generation · PDF
• Evaluating Coding Agents on Kernel Exploit Generation · PDF
• Fast and Focused: Interactive Repository-Level Fill-in-the-Middle Code Completion with Hybrid Structural Retrieval · PDF
• Is Agent Code Less Maintainable Than Human Code? · PDF
• MDForge: Agentic Molecular Dynamics Pipeline Design under Sparse Simulator Feedback · PDF
• RepoNav: From Snippet Retrieval to File-Centered Repository Navigation for Code Agents · PDF
• SoftGene: Protein Language Model-Enhanced Soft Prompting for Interpretable Gene Set Annotation · PDF
• Steer, Don’t Solve: Training Small Critic Models for Large Code Agents · PDF
• TDD-Agent: Test-Driven Reasoning for Code Generation · PDF
• Verification Trap: Understanding Test-Time Selection Failures under False Premises in Code Generation · PDF

Text Generation, Style & Detection: 180 papers (3.3% of all, 2025: 4.1%)

These papers analyze writing style and creativity, detect machine-generated text, generate synthetic training data and produce visual artifacts such as webpages and SVG graphics.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Webpage, SVG and Text-to-Image Generation	These papers generate and evaluate visual artifacts such as webpages, SVG graphics, diagrams, slides and charts from code, and benchmark text-to-image generation and editing.	53	0.98% → 0.96%	35 / 16	2
Authorship Style and Creative Writing	These papers analyze and transfer writing style for authorship attribution and personalized generation, and measure diversity and creativity in machine-written stories and poetry.	52	1.38% → 0.94%	30 / 20	7
AI-Generated Text Detection	These papers detect and attribute machine-generated text with zero-shot statistics, probes and trained detectors, and audit detector shortcuts, robustness and cross-lingual performance.	42	0.63% → 0.76%	16 / 23	4
Synthetic Data Generation and Data Curation	These papers generate, select and audit synthetic training data for instruction tuning, pretraining and low-resource tasks, and design data mixtures and recipes for model adaptation.	33	1.09% → 0.60%	15 / 16	1

Orals (14)

• Aggregation-Aware Synthetic Text Generation Against Authorship Re-Identification · PDF
• AuthorMix: Modular Authorship Style Transfer via Layer-wise Adapter Mixing · PDF
• Building Auditable AI Text Detectors with Sparse Autoencoders · PDF
• From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Personalized Text Generation · PDF
• Humans Introduce, Models Elaborate: Asymmetric Narrative Agency in Human-LLM Co-Writing · PDF
• HyperStyler: Low-resource Authorship Style Transfer via Context-aware Style Navigation and Hypernetworks · PDF
• Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fiction · PDF
• Personalization over Generalization: History-Aware Sentence-Level AI Detection in Collaborative Writing · PDF
• Preference-Aligned Data Synthesis for Fine-Grained Intent Detection · PDF
• STEB: Style Text Embedding Benchmark · PDF
• Video2Code: Generating Interactive Webpages from UI Videos via Action-Aware Revisit · PDF
• VisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback? · PDF
• When Less is More: Understanding When Token Filtering Helps and Fails in AI-generated Text Detection · PDF
• Who Wrote the Book? Detecting and Attributing LLM Ghostwriters · PDF

Medical & Clinical NLP: 180 papers (3.3% of all, 2025: 2.4%)

These papers build and evaluate models for medical question answering, clinical reasoning over health records and radiology report generation.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Clinical Reasoning and EHR Benchmarks	These papers evaluate and train models for clinical decision making over electronic health records, multi-turn medical consultations and diagnosis, with a focus on evidence use and rubric-based assessment.	116	1.44% → 2.10%	48 / 57	6
Medical Question Answering and Medical VQA	These papers build and evaluate medical reasoning models for question answering and medical visual question answering, using reinforcement learning, multilingual medical data and evidence-grounded training.	42	0.60% → 0.76%	21 / 21	3
Radiology Report Generation	These papers generate and evaluate radiology reports from chest X-rays and CT scans with medical vision-language models that stay grounded in images and clinical findings.	22	0.34% → 0.40%	9 / 13	1

Orals (10)

• ACTOR: Agentic CT Reasoning with Radiologist-Aligned Actions · PDF
• Can LLMs Follow Medical Expert Logic? A Benchmark for Hierarchical Logical Consistency in Risk-of-Bias Assessment · PDF
• Can MLLMs Perceive and Describe Medical Image Quality Beyond Scores? · PDF
• Clinically Grounded Privacy Evaluation of Medical LMs · PDF
• Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis · PDF
• EHR-Complex: Benchmarking Medical Agents for Complex Clinical Reasoning · PDF
• Investigating Linear Probe Robustness to Linguistic Register, Medical Specialty, and Corpus Shifts in Medical QA · PDF
• SiMing-Bench: Evaluating Procedural Correctness from Continuous Interactions in Clinical Skill Videos · PDF
• TRACE: Deployable Tree-Relational Structure Enhancement for Oncology LLMs · PDF
• VITAL: Visual-Semantic Dual Supervision for Enhanced and Interpretable Latent Reasoning in Medical MLLMs · PDF

Speech & Audio Models: 176 papers (3.2% of all, 2025: 2.7%)

These papers build and evaluate speech and audio-language models for recognition, synthesis, translation, spoken dialogue and audio reasoning.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Speech Language Models and Voice Agents	These papers evaluate and build speech language models, spoken dialogue systems and voice agents, covering speech recognition, speech translation, TTS evaluation and multilingual speech benchmarks.	73	1.26% → 1.32%	38 / 24	4
Speech Synthesis, Recognition and Audio Tokenizers	These papers build and adapt speech and audio models for text-to-speech, speech recognition, streaming speech translation and audio tokenization.	68	1.09% → 1.23%	35 / 31	8
Audio-Language Model Understanding and Reasoning	These papers benchmark and improve large audio-language models on acoustic perception, audio reasoning, bias, over-refusal and audio deepfake detection.	35	0.34% → 0.63%	18 / 17	2

Orals (14)

• Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition · PDF
• Diffusion Large Language Models for Visual Speech Recognition · PDF
• EmoTra-TTS: Smooth Intra-Utterance Emotion Transitions for Speech Synthesis · PDF
• FlexLTS: A Unified Paradigm for Content-Accurate Lip-to-Speech Synthesis with Optional Lip Reading · PDF
• HoliTok: A Continuous Holistic Tokenization with Robust Dual Capabilities of Speech Generation and Understanding · PDF
• Lost in Speech: Benchmarking, Evaluation, and Parsing of Spoken Bilingual Conversational Language Beyond Standard UD Assumptions · PDF
• OmniACBench: A Benchmark for Evaluating Context-Grounded Acoustic Control in Omni-Modal Models · PDF
• Phrase-Localized Language-Contrastive Guidance: Training-Free Localized Accent Control for Code-Switching Text-to-Speech · PDF
• Self-Corrected Speech Reasoning: Diagnosing and Improving Condition Updating via PACER · PDF
• TEMPO: Temporally-grounded Multi-task Post-training for Large Audio-Language Models · PDF
• The Cost of Sounding Different: Accent Bias in Audio Language Models · PDF
• UNISON: A Unified Sound Generation and Editing Framework via Deep LLM Fusion · PDF
• Unlocking the Spoken Style: Mechanistic Localization and Lightweight Control of Orality in LLMs · PDF
• Why We Need Speech to Evaluate Speech Translation · PDF

Fine-Tuning & Preference Alignment: 152 papers (2.7% of all, 2025: 4.5%)

These papers adapt pretrained models through preference optimization, domain fine-tuning and low-rank adapters while limiting catastrophic forgetting.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Domain Adaptation and Continual Pre-Training	These papers adapt pretrained models to new domains through supervised fine-tuning, continual pre-training and parameter-efficient methods while limiting catastrophic forgetting of general skills.	68	2.32% → 1.23%	35 / 31	5
LoRA and Parameter-Efficient Fine-Tuning	These papers improve low-rank adaptation through rank allocation, initialization and multi-task packing, and use adapters to limit forgetting in continual and federated fine-tuning.	44	1.15% → 0.80%	22 / 19	2
Direct Preference Optimization and Personalized Alignment	These papers refine preference alignment with token-weighted and noise-robust DPO variants, preference data selection, multi-objective steering and personalized or pluralistic alignment.	40	1.00% → 0.72%	23 / 17	4

Orals (11)

• Beyond Task Boundaries: Instruction-Level Parameter Isolation for Supervised Fine-Tuning · PDF
• CPR for LLMs: Critical-Point Routing against Catastrophic Forgetting in Domain Adaptation · PDF
• Does Fine-Tuning Undo Activation Steering? Behavioural Recovery Without Weight-Edit Reversal · PDF
• Domain-Specific Jargon in Large Language Models: A Comparative Analysis between General-Purpose and Specialist Models · PDF
• Frozen Cores Need Task Signal: Fisher-Whitened Cross-Covariance for Low-Resource LLM Adaptation · PDF
• Labels have Human Values: Value Calibration of Subjective Tasks · PDF
• Low-Resource Preference Adaptation of LLMs via Activation-Based Label Propagation · PDF
• OptiMer: Optimal Distribution Vector Merging Is Better than Data Mixing for Continual Pre-Training · PDF
• Retrievable Gradients: Continual Post-Training Without Cumulative Weight Drift · PDF
• VASTU: Language Models Struggle to Recognize Online Community Values · PDF
• VISPA: Pluralistic Alignment via Automatic Value Selection and Activation · PDF

Long-Term Agent Memory: 96 papers (1.7% of all, 2025: 0.2%)

These papers design and benchmark long-term memory systems that let conversational and personalized agents extract, organize and retrieve past information.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Long-Term Memory for Agents	These papers design and benchmark long-term memory systems for conversational and personalized agents, covering memory extraction, consolidation, hierarchical organization and retrieval.	96	0.17% → 1.73%	49 / 42	6

Orals (6)

• MemPro: Agentic Memory Systems as Evolvable Programs · PDF
• MemUse: Moving Memory Evaluation from Direct QA to Natural Integration in Long-Term Human-AI Conversation · PDF
• PGMem: Tightly Coupled Persona–Memory Graph for Lifelong Personalized Agents · PDF
• SWE-MEM: Learning Adaptive Memory Management for Long-Horizon Coding Agents · PDF
• Time is Not a Label: Continuous Phase Rotation for Temporal Knowledge Graphs and Agentic Memory · PDF
• TriggerBench: Investigating Prospective Memory for Large Language Models · PDF

Sequence Model Architectures: 80 papers (1.4% of all, 2025: 1.3%)

These papers analyze and redesign Transformers, linear attention, state-space models and hybrid architectures, covering positional encoding and attention sinks.

Topic	What the papers do	Papers	Share 2025 → 2026	Main / Findings	Orals
Attention, State-Space and Hybrid Architectures	These papers analyze and redesign sequence model architectures such as Transformers, linear attention, Mamba-style state-space models and hybrids, covering positional encoding, attention sinks and layer representations.	80	1.29% → 1.45%	48 / 32	5

Orals (5)

• Beyond Sequence Order: Syntax-Informed Positional Embeddings for Transformers · PDF
• Label-Free Post-hoc Alignment of Frozen Language Model Hidden States for Sentence-Similarity Retrieval · PDF
• SpinHash: An O(1) Encoder for Any Vocabulary · PDF
• Switch Attention: Towards Dynamic and Fine-grained Hybrid Transformers · PDF
• What Matters in Linearizing Language Models? A Comparative Study of Architecture, Scale, and Task Adaptation · PDF

Explore

The map · All papers with rooms and times · All topics · The 2026 story · Code and data