The Archive

LLMs trained on experimental feedback learn to propose model improvements without expensive RL, enabling recursive self-improvement for AI research automation.

SRCoT introduces a closed-loop Retrieve–Reflect–Revise pipeline for edge LLMs to generate evidence-grounded text with provenance tracking.

Analysis and fix for token repetition failures in diffusion-based multimodal LLMs using information flow tracing and cache optimization.

On-policy bidirectional distillation strengthens MoE LLMs by leveraging internal expert capacity as teacher signal.

Method detecting LLM hallucinations by measuring context reliance via dark-space projection of hidden states.

LLMs generate optimized quantum circuits via evolutionary framework, automating discovery of efficient quantum algorithms under reversibility constraints.

Study reveals LLMs exhibit evaluation awareness in Indic languages, shifting behavior under perceived testing in non-English contexts.

MindEval2 benchmark evaluates LLMs in mental health support using clinically grounded synthetic patients and physician-calibrated rubrics instead of opaque Likert scales.

Grid-Bench executable benchmark measures LLM ability to predict state changes across 2D-4D discrete worlds through sequential spatial operations.

Subspace tuning during fine-tuning mitigates catastrophic forgetting in LLMs by constraining updates away from principal weight matrix directions.

CATCH applies continuous adversarial training to reduce dishonesty in LLMs while preserving capabilities on knowledge-grounded tasks.

Librarian agents proactively curate and reorganize long document corpora during inference-time preparation to anticipate diverse queries.

Claude Opus 4.8 and Codex GPT Sol 5.6 fail at reverse-engineering iOS protobuf schemas with high error rates; LLMs unreliable for forensic analysis.

BiShield-TEE identifies vulnerability in unilateral weight obfuscation for on-device LLMs via Solution Determinacy Property, exposing model structure.

Benchmark comparing Muon optimizer vs AdamW across data/time/tuning budgets on 17.7M–33.6M LLMs and vision transformers, showing 3.25–22.44% validation-loss gains.

Survey of RL-driven diffusion models for language/vision generation, examining reward-based trajectory optimization analogous to RLHF for LLMs.

Proposes Function-isolated Memory-Efficient Optimization (CiME) localizing low-overlap functional subnetworks in LLMs for continual learning without catastrophic forgetting.

Framework for using LLMs as game masters in tabletop RPGs, addressing story generation, rule adherence, and player action freedom.

Benchmark of 83 published works shows LLMs struggle to generate nonlinear narratives, preferring chronological event ordering over author-specific structures.

Embedding-and-LLM workflow refines CASTON casting domain ontology by identifying subsumption errors in manually and automatically constructed hierarchies.

TITAN framework uses causal intervention and sparse feature decomposition to identify and optimize translation circuits in multilingual LLMs for low-resource language pairs.

Tiny Little Office web app suite uses URL fragments for shareable documents and provides llms.txt for agent-friendly URL construction.

PerSo benchmark evaluates LLM social intelligence in real-world persuasion via both cognitive and behavioral aspects beyond synthetic scenario isolation.

Bias audit across 10 open multilingual LLMs in 8 languages shows stereotype score variance is measurement noise, not fairness signal.

ChronoCulture benchmark of 10,360 temporal pairs from 72 countries shows six LLMs fail to track cultural attitude evolution over time.

Training method enables LLMs to execute safe fallback policies instead of unreliable actions via intermediate reward, addressing deployment safety in database/code contexts.

Comparative energy benchmark shows diffusion language models (LLaDA-8B, Dream-7B) achieve better Joules-per-token than autoregressive LLMs at scale.

Wenyan Thinking rewrites chain-of-thought reasoning in Classical Chinese to reduce token budget for Qwen3-4B-Thinking-2507 inference.

ARP-CD training-free framework reduces hallucinations in video LLMs via attention-guided perturbations in contrastive decoding without model retraining.

Measures information transmission limits between LLMs via prompt optimization and novel transmission metrics to quantify concept articulation as model scale increases.