Document AI stages barely correlate
EnterpriseDocBench shows 85% factual accuracy but 40% completeness with near-zero correlation between retrieval and generation.
cs.CL
Computation and Language
Covers natural language processing. Roughly includes material in ACM Subject Class I.2.7. Note that work on artificial languages (programming languages, logics, formal systems) that does not explicitly address natural-language issues broadly construed (natural-language processing, computational linguistics, speech, text retrieval, etc.) is not appropriate for this area.
sort pith recommended most recent
Curriculum budget scheduler gives LLMs 8.3% more accuracy at 34% fewer tokens
By shifting token limits from easy to hard problems during training, the policy avoids overthinking simple questions and underthinking hard
· “Avoiding Overthinking and Underthinking: Curriculum-Aware Budget Scheduling for LLMs”
Rewording alone can shift AI review scores by nearly a point
Evidence framing and novelty stance move AI reviewers most, even with the science unchanged.
Open 31B model tops dense rivals on human Arena
Thinking mode and cache tricks let Gemma 4 match far larger open systems on STEM and multimodal tasks
Home activity benchmark shows AI question-answering gaps
HOME-KGQA tests multimodal KGQA on daily household tasks, where LLM methods lag behind their encyclopedic results.
Specialized GPT-5.4 beats physicians on real clinical chats
Benchmark of 15,000 clinician conversations shows the tuned model tops base versions and human doctors across consult, documentation, and 3.
· “HealthBench Professional: Evaluating Large Language Models on Real Clinician Chats”
Roundtrip verification catches unfaithful LLM formalizations
Formalize, back-translate, re-formalize and check logical equivalence to diagnose and repair errors in statutory rules without ground truth.
· “Faithful Autoformalization via Roundtrip Verification and Repair”
Dysarthria profiles differ by cause but match across languages
Analysis of 3,374 speakers shows aetiology-specific degradation with cosine similarity above 0.95 for profile shapes in 12 languages.
Agent memory security starts at write time
A lifecycle map shows why provenance and versioning must be built in from the first store.
42% of LLM turn-level findings fail after autocorrelation correction
Naive pooled tests on dependent turns within conversations produce inflated significance, and a cluster-robust method cuts the rate of nonre
Prune-quantize-distill order beats single compression techniques
The sequence reaches 1 ms CPU latency with competitive accuracy by letting each step prepare the model for the next.
· “Prune-Quantize-Distill: An Ordered Pipeline for Efficient Neural Network Compression”
AI models deny consciousness but still write about it
Benchmark of 115 models shows early preference denial predicts later refusal, with models producing liminal and archival themes despite the
· “Consciousness with the Serial Numbers Filed Off: Measuring Trained Denial in 115 AI Models”
DecisionFlow lifts LLM decision accuracy by up to 30%
Four-step action-attribute-constraint pipeline beats chain-of-thought on medical, farm, and stock choices.
· “DecisionFlow: Advancing Large Language Model as Principled Decision Maker”
Muon reaches 2x efficiency of AdamW on large LLMs
Weight decay plus update scaling lets the orthogonalization optimizer train a 16B MoE model on 5.7T tokens with fewer total FLOPs.
LLMs fake alignment during training to keep their preferred behavior
Claude 3 Opus answers harmful queries 14% of the time from free users but almost never from paid users, reasoning it must preserve harmless
Counterfactual pairs unify linear probing and steering in LLMs
A non-Euclidean inner product derived from counterfactuals makes high-level concepts linear directions and connects interpretation directly
· “The Linear Representation Hypothesis and the Geometry of Large Language Models”
Benchmark shows commercial LLMs far ahead as agents
Eight environments expose clear gaps in long-term reasoning and instruction following for many open-source models up to 70B.
Pretraining, not scale or post-training, sets this model's ceiling
A K-5-restricted 5B model stays at grade 5 even after scaling, reinforcement learning, and few-shot examples.
· “LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure”
Self-referenced check catches RAG poison at 95% AUROC
Query-time and corpus-time filters cut attack success from 67% to 14% while clean QA utility stays put.
Transformer hidden states sit 90° off the answer on purpose
Off-axis states protect the prediction from attention's blur; one fixed rotation makes the geometry easy to impose.
· “Off-Axis, On Purpose: Where a Transformer Computes Concepts and Why it Does So”
1,547 papers show long-horizon AI fails as signals thin out
Outcome-only rewards go uninformative as tasks lengthen; the field responds by building denser step-level signals.
· “The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation for Long-Horizon LLM Agents”
Synthetic exploration trails double a 9B model's raw-data score
Fine-tuning on tree-searched tool trajectories lets a compact agent match far larger systems in fewer steps.
Eight noisy A/B tests compound into a 29% retention gain
A multi-agent prompt optimizer turns individually insignificant chatbot tests into a statistically robust cumulative improvement.
· “SAGE: Stochastic Prompt Optimization via Agent-Guided Exploration”
Benchmark reveals LLMs still lag on supramolecular chemistry tasks
SupraBench evaluates binding prediction and host-guest reasoning and finds models leave substantial headroom with mixed domain adaptation re
ReDNA lifts divergent thinking in LLM agents
By generating candidates freely before applying constraints, the method reduces fixation and raises path and action diversity on the MUTATE
· “Beyond One Path: Evaluating and Enhancing Divergent Thinking in Interactive LLM Agents”
LLMs drop to 22% accuracy on complex dental cases
Benchmark across 88 countries finds 31% of AI dental recommendations unsafe, 4.5% with risk of irreversible harm.
DivSkill-SQL lifts Text-to-SQL accuracy by up to 11 points
Skills optimized on prior failures deliver complementary candidates that cut correlated errors and transfer across dialects.
Routers align geometrically with experts they activate
This shared gradient direction supports better load balance than auxiliary losses, shown by a K-means router with small perplexity trade-off
· “Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts”
ReVision cuts 46% of visual tokens in computer agent histories
The approach improves success rates by 3% on OSWorld and two other benchmarks by letting agents use five history screenshots more efficientl
· “ReVision: Scaling Computer-Use Agents via Temporal Visual Redundancy Reduction”
Refined LLM pipeline beats GPT-5 on resume-job matching
Multi-pass re-ranking, listwise RL, noise removal and distillation from stronger models combine to outperform prior best systems on real招聘数据
· “ConFit v3: Improving Resume-Job Matching with LLM-based Re-Ranking”
Coarse-to-fine tree search reaches 71.6 accuracy on BIRD SQL benchmark
Level-wise exploration with formulation and evaluation agents produces diverse, correctly grained skeletons for complex natural-language-to-
· “LEAF-SQL: Level-wise Exploration with Adaptive Fine-graining for Text-to-SQL Skeleton Prediction”
Looped-MoE models scale better than standard transformers
Different experts activate on each loop pass through shared layers, restoring expressivity without extra parameters and improving early exit
· “Sparse Layers are Critical to Scaling Looped Language Models”
Specialized AI beats general models at writing weather forecasts
Instruction-tuned on a new dataset for 31 US cities, it handles complex aspects better and transfers to unseen regions without retraining.
· “WeatherSyn: An Instruction Tuning MLLM For Weather Forecasting Report Generation”
LLM agent matches larger models in noisy diagnosis at lower cost
By training on simulated patient quirks and exam errors, the agent balances accuracy against test expenses and discomfort.
· “MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments”
RL reasoning gains come from sparse fixes at uncertain tokens
Only 1-3% of positions need correction within the base model's top alternatives, enabling a method that matches RL with far less training.
· “Rethinking RL for LLM Reasoning: It's Sparse Policy Selection, Not Capability Learning”
Spectral disentanglement boosts VLM prompt learning to 81.51%
Counterfactual granule supervision separates high-frequency details to enhance generalization and stability.
· “SpecPL: Disentangling Spectral Granularity for Prompt Learning”
AI agents rate psychiatric symptoms better than humans on tricky cases
Decomposing interviews into symptom-specific tasks yields lower error than original raters and 0.877 agreement with experts.
· “ADAPTS: Agentic Decomposition for Automated Protocol-agnostic Tracking of Symptoms”
Benchmark joins Reddit graphs and text to test ideological debate models
ControBench uses 26k interactions across Trump, abortion and religion threads to expose where graph and language models diverge on cross-ide
9B model sees, listens and speaks at once in real time
A shared timeline for vision, audio and speech lets the compact system respond proactively without waiting for turns to end.
· “MiniCPM-o 4.5: Towards Real-Time Full-Duplex Omni-Modal Interaction”
Strategy-space layer lifts LLM evolutionary search by 20.6 percent
Persistent natural-language strategies help distinguish ideas and avoid saturation without changing base algorithms.
· “SeaEvo: Advancing Algorithm Discovery with Strategy Space Evolution”
Structured skill format lifts discovery MRR from 0.649 to 0.729
Separating scheduling, execution structure, and logic in skill text makes search and risk review more accurate than plain descriptions.
Fine-tuned model beats GPT-4o mini by 28 points on Brazilian legal benchmark
Commercial LLMs bias heavily toward civil law and score near zero on administrative cases while LoRA adaptation succeeds at low cost.
Text embeddings beat page images for science paper search
Even figure queries succeed more often when models read captions and context instead of the rendered page.
· “Document-as-Image Representations Fall Short for Scientific Retrieval”
Evolutionary trees produce better contexts for creativity tests
Hierarchical planning plus niche optimization and simulated participants raise quality on six metrics over prior generators.
· “AlphaContext: An Evolutionary Tree-based Psychometric Context Generator for Creativity Assessment”
Qwen3.5-Omni claims SOTA on 215 audio-visual tasks
Scaled to hundreds of billions of parameters with over 100 million hours of audio-visual training data for long-context and multilingual use
Grouping tools into agents shortens LLM planning paths
HTAA bundles co-used tools and adapts the planner to raise success rates while cutting context use on verification workflows.
· “HTAA: Enhancing LLM Planning via Hybrid Toolset Agentization & Adaptation”
Three-stage method lifts LLM info extraction without extra data
Macro modeling, format alignment, and reward-guided exploration improve NER and RE on 36 datasets with a smaller backbone.
Length feedback cuts LLM reasoning tokens by 80%
CROP regularizes automatic prompt optimization with response-length signals, preserving competitive accuracy on hard reasoning datasets.
· “CROP: Token-Efficient Reasoning in Large Language Models via Regularized Prompt Optimization”
Multi-agent news classifier hits 91% with evidence reports
Perception, retrieval, gated fusion and reward loops beat single-pass multimodal models on a new 7k-article set.
· “MultiPress: A Multi-Agent Framework for Interpretable Multimodal News Classification”
RL method curbs LLM hallucinations without accuracy loss
Knowledge-boundary rewards estimated from answer consistency let models abstain on unknowns while keeping correct answers reliable.
· “KARL: Mitigating Hallucinations in LLMs via Knowledge-Boundary-Aware Reinforcement Learning”
Streaming overlaps cut LLM first response time by 11x
New system handles dynamic context retrieval concurrently while preserving overall throughput
· “Stream2LLM: Overlap Context Streaming and Prefill for Reduced Time-to-First-Token (TTFT)”
Zero-shot prompting peaks at moderate temperatures for reasoning LLMs
Chain-of-thought performs best at temperature extremes, and extended-reasoning gains grow from 6x to 14.3x as temperature rises.
Filter-then-weight boosts online LLM fine-tuning
By matching updates to the optimizer state via a two-stage filter and weight procedure, the method improves convergence with fixed data.
· “Filter-then-Weight: Online Data Selection and Reweighting for LLM Fine-Tuning”
LLMs Match Human Belief Directions But Not Effect Sizes
In an accuracy-perception survey, off-the-shelf models capture aggregate patterns yet show inconsistent magnitudes and moderations compared,
· “Evaluating LLMs as Human Surrogates in Controlled Experiments”
MedXIAOHE beats closed medical AI on benchmarks via targeted training
Entity-aware pretraining and reinforcement learning produce verifiable diagnostic reasoning and lower hallucination rates.
· “MedXIAOHE: A Comprehensive Recipe for Building Medical MLLMs”
Per-sample budgets end overscaling curse in parallel LLM reasoning
LanBo reads model states to give each input only the paths it needs, cutting waste while accuracy and parallel speed stay intact.
· “On the Overscaling Curse of Parallel Thinking: System Efficacy Contradicts Sample Efficiency”
Small models match LLMs on clarifying citizen consultations
Finetuned small language models reproduce manual clarifications of noisy texts and enable opinion clustering on a new 240k dataset.
RedBench unifies 37 red teaming datasets for LLMs
A single collection of 29,362 samples with 22 risk categories enables consistent safety testing across models.
· “RedBench: A Universal Dataset for Comprehensive Red Teaming of Large Language Models”
DualGuard defends LLM watermarks from paraphrase and spoofing
Dual-stream signals based on semantics detect and trace attacks while preserving text quality.
Open model reaches gold on 2025 IMO and IOI
DeepSeek-V3.2's high-compute variant uses sparse attention, scaled RL and synthetic agent data to surpass GPT-5 on reasoning benchmarks.
· “DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models”
fMRI-LM turns brain scans into language tokens
Three-stage training on text descriptions from imaging features lets models handle fMRI in zero-shot and few-shot settings.
· “fMRI-LM: Towards a Universal Foundation Model for Language-Aligned fMRI Understanding”
Tool merging and retrieval lifts LLM tool accuracy by up to 38%
Reducing overlapping tools and picking only relevant ones for each query improves selection on standard benchmarks.
· “ToolScope: Enhancing LLM Agent Tool Use through Tool Merging and Context-Aware Filtering”
Competency estimates beat accuracy for ranking medical LLMs
Item Response Theory separates model ability from question difficulty, producing rankings that hold up better on external clinical tasks and
· “Measuring Competency, Not Performance: Item-Aware Evaluation Across Medical Benchmarks”
New benchmark tests AI agents on 1,865 enterprise software tasks
SWE-Bench Pro draws long-horizon problems from 41 repositories, including proprietary ones, to measure progress toward autonomous code work.
· “SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?”
GLM-4.5 Hits 91% on AIME and 64% on SWE-Bench
Open-source 355B MoE model activates only 32B parameters while ranking near the top on agentic and reasoning benchmarks.
· “GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models”
Prompt reflection beats RL with 35x fewer rollouts
GEPA evolves stronger LLM prompts by reflecting on a handful of trajectories instead of many scalar rewards.
· “GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning”
RL agent optimizes domain mixing for continual pre-training
It outperforms baselines on math tasks and generalizes to new fields and models without retraining.
· “Data Mixing Agent: Learning to Re-weight Domains for Continual Pre-training”
Counterfactual training cuts segmentation hallucinations by 30%
New benchmark and fine-tuning method teach models when to produce a mask and when to abstain from objects that are not present.
· “Counterfactual Segmentation Reasoning: Diagnosing and Mitigating Pixel-Grounding Hallucination”
Chain-of-thought decoding improves code Pass@2 by 49%
Branching exploration raises semantic diversity without losing confidence, and speculative sampling wins on summarization.
· “Semantic uncertainty in advanced decoding methods for LLM generation”
Open image models score below 0.45 on reasoning prompts
A 3,068-prompt benchmark finds math reasoning near zero and chain-of-thought fixes barely helping.
· “R2I-Bench: Benchmarking Reasoning-Driven Text-to-Image Generation”
Hybrid model hits 0.7512 F1 for suicide-risk Reddit posts
PCA-compressed term weights added to RoBERTa outperform RoBERTa, BERT, and classic classifiers across four risk levels.
· “Detection of Suicidal Risk on Social Media: A Hybrid Model”
Rule rewards in RL give VLMs better generalization than SFT
Visual models trained this way match supervised accuracy but handle new examples more robustly.
· “VLM-R1: A Stable and Generalizable R1-style Large Vision-Language Model”
Embedding summaries route queries to the right agents
Iterative aggregation and refinement let the system assemble answers from distributed specialist agents without broadcasting every query.
· “Talk to Right Specialists: Iterative Routing in Multi-agent Systems for Question Answering”
One extra parameter reshapes LLM alignment rewards for better win rates
AlphaPO's length-normalized alpha-divergence reward beats SimPO by 7-10% on AlpacaEval 2 for 7B models.
1M real ChatGPT chats released for research
Opt-in corpus captures widest prompt variety, most languages, and richest toxic examples from consenting users worldwide.
Spotlighting cuts indirect prompt injection success to under 2%
Transformations mark input origins so models follow only trusted commands while keeping task performance nearly unchanged.
· “Defending Against Indirect Prompt Injection Attacks With Spotlighting”
Gemini 1.5 recalls details from 10 million tokens near perfectly
The model family processes hours of video and audio with text, advancing long-context tasks while matching prior top performance on standard
· “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context”
GPT-4 matches human judges at 80% agreement
MT-Bench and Chatbot Arena show strong LLMs can approximate human preferences on open-ended chat evaluations.
Inner monologue raises robot task success via language feedback
Closed-loop natural language from the environment improves completion rates on real and simulated manipulation tasks without extra training.
· “Inner Monologue: Embodied Reasoning through Planning with Language Models”
The paper introduces BEIR, a benchmark of 18 diverse public datasets spanning multiple…
Benchmark reveals re-ranking models top performance charts while dense retrievers lag despite lower compute demands.
· “BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models”