Streamliners for Answer Set Programming
StreamLLM approach adapted to Answer Set Programming: LLMs generate candidate streamliner constraints to reduce combinatorial search space.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
StreamLLM approach adapted to Answer Set Programming: LLMs generate candidate streamliner constraints to reduce combinatorial search space.
Multi-turn dialogue study reveals LLMs exhibit divergent repair behaviors: some resist user corrections, others highly susceptible to manipulation.
As LLMs transition from simple text generation to complex reasoning, reinforcement learning (RL) plays a central role. Algorithms like Group Relative Policy... As LLMs transition from simple text generation to complex reasoning, reinforcement learning (RL) plays a central role…
The long-term risks of the All-In Podcast, illustrated. | Image: Cath Virginia / The Verge, Turbosquid, Getty Images One of the most mortifying things about knowing a lot of techies is listening to them tell me excitedly about some very important discovery that they believe…
MathNet: 30.7K multilingual Olympiad problems across 47 countries benchmarking math reasoning in LLMs and retrieval in embeddings.
Empirical study: LLMs generalize under weak supervision for reasoning via RLVR when training reaches reward saturation.
System Dynamics benchmarks (CLD, Discussion Leaderboards): cloud LLMs 77–89% vs. best local 77% on causal diagram extraction.
GSQ applies Gumbel-Softmax sampling to scalar quantization, achieving <4bpp accuracy without vector-quantization complexity for LLM deployment.
Mechanistic analysis of three jailbreak methods (SFT, RLVR, abliteration) on open-weight LLMs; reveals divergent behavioral and internal failure modes.
DPO-based alignment framework instills musical constraints in LLMs for lyric-to-melody generation without annotation.
NI Sampling optimizes token ordering in discrete diffusion LLMs to accelerate parallel decoding beyond heuristic step-selection baselines.
DESPITE benchmark (12K tasks) reveals LLMs as robotic planners generate dangerous plans 28.3% of the time despite strong planning ability; scale ≠ safety.
Framework for dynamic mid-generation abstention in LLMs using principled decision rules to early-terminate unpromising reasoning traces.
DORA Explorer improves LLM agent exploration diversity in sequential decision-making without training via entropy-based sampling.
Introduces safety token regularization to preserve alignment properties during domain-specific fine-tuning of LLMs.
New CogBiasESC dataset for training LLMs to detect and address cognitive distortions in emotional support conversations.
LLMs generate executable legal decision models from Dutch environmental law text using structured intermediate representations, evaluated on production Omgevingsloket platform.
Identifies vulnerability in majority-voting multi-agent LLM systems against prompt injection, proposes token-level collaboration for robustness.
CGM-Agent enables privacy-preserving QA over continuous glucose monitor data using LLMs with local computation to protect sensitive health records.
Adaptive Contrastive Decoding mitigates over-refusal in safety-aligned LLMs by keeping non-refusal tokens in candidate lists while maintaining high refusal for malicious queries.
Frontier LLMs fail citation accuracy for rare diseases; HEG-TKG grounds clinical claims in temporal knowledge graphs with evidence traceability to resolve provenance gap.
Cross-model semantic disagreement complements self-consistency for uncertainty quantification in LLMs, detecting overconfident errors when single-model consensus fails.
DeepInsightTheorem framework trains LLMs for informal theorem proving by teaching recognition of core problem-solving techniques via hierarchical reasoning.
Cross-linguistic study of politeness effects on 5 LLMs (Gemini-Pro, GPT-4o Mini, Claude 3 Sonnet, DeepSeek-Chat, Llama 3) via 22,500 English/Hindi/Spanish prompts.
Dual-aspect evaluation framework for 4 LLMs (GPT-4o, Claude 3 Opus, Gemini 1.5 Pro, Grok-1) on Vietnamese legal text simplification: accuracy, readability, consistency.
BAGEL benchmark evaluates specialized animal knowledge in LLMs using bioRxiv, biotic interaction, audio, and Wikipedia sources with closed-book QA pairs.
Benchmark compares token pruning compression across Qwen3, Gemma-3, Llama-3, Aya for Korean-centric NLP with English-Korean vocabulary optimization.
Conformal prediction framework for LLMs using Layer-Wise Information scores from internal representations for robust uncertainty calibration.
RISE: scalable data attribution method for LLMs focusing on output-layer influence hotspots via sketching instead of full gradient indexing.
AtManRL trains differentiable attention masks via RL to improve faithfulness of chain-of-thought reasoning in LLMs by identifying decision-critical tokens.