The Archive

Framework for controlling LLM post-training behavioral sharpening via RL-based pass@k optimization adapted to policy gradients.

DepthGraft uses cross-layer KV reconstruction as structural regularization to improve information flow in deep LLM architectures.

SCOPE framework trains LLM web agents via progressive skill acquisition, decomposing long-horizon tasks into learnable sub-skills with sparse rewards.

Study shows preprocessing decisions in fine-tuning data significantly impact fairness and utility in LLM-based recommendation systems.

LUCA synthetic benchmark evaluates LLM agents on experimental reasoning, diagnosis, and intervention across seven frontier models with outcome-independent metrics.

Experimental study comparing hybrid retrieval, NLI models, and open-source LLMs for scientific claim verification using conservative fusion.

PACERec progressively adapts LLM item embeddings for sequential recommendation via collaborative evidence transfer to long-tail items.

PRIME framework routes LLM scientific idea revisions by predicting intervention effects, addressing feedback-driven iteration mismatch.

LLM pipeline extracting structured criminal court data via iterative attribute exploration for empirical legal research.

VERS steers reasoning LLMs by editing natural language explanations of activations at doubt points, improving GSM8K accuracy.

Q-CamTuning, a multimodal LLM for fine-grained image quality assessment and camera tuning in livestreaming applications.

BLPO improves long-horizon LLM agent training via cross-episode advantage pooling for better credit assignment.

BERT-large fine-tuned on Gemma 4 annotations achieves 0.589 Kendall's τ for ordinal literary sentiment analysis.

CAREFUL-Bench extends function calling evaluation to detect unpenalized destructive side effects in multi-turn LLM interactions.

SHq learns per-bit-width lookup-table grids for LLM quantization, outperforming fixed NF4 and BOF4 grids.

Framework decomposes LLM predictive uncertainty to isolate impact of underspecified inputs vs. clarifications on outcomes.

Conjecture-and-refutation verification method catches misaligned LLM agent actions that evade standard confirmation checks.

PlagiaCorpus: 198K Arabic-English dataset for detecting LLM-generated plagiarism and cross-lingual reuse across five difficulty levels.

Analysis of 150K LLM conversations shows political ideology and user attributes shape LLM receptiveness and speaker validation in debates.

Iluvatar CoreX Tiangai 300 outperforms Nvidia H200 in select LLM inference, long-context, and MoE benchmarks.

Kyle Kingsbury warns LLMs undermine trust in software docs and quality signals, prompting debate on verification and credibility for maintainers.

MIRA applies mutual information and Gaussian perturbations for source-free test-time adaptation, addressing LLM overconfidence under distribution shift.

TrackMem enhances LLM agent memory for financial dialogue by tracking belief evolution and value supersession rather than retrieval similarity.

RubricWorld optimizes LLM grading rubrics using world models to predict edit effects, reducing validation rescoring labor.

Supply-chain attack vector in LLM agents: tool code updates bypass model-level defenses via context drift.

Hybrid narrative framework balances static content with real-time LLM generation for interactive storytelling coherence.

RECEIPT identifies and repairs reliability failures when LLM agents call web services with unbounded tool-call returns.

LLM-guided active learning for protein prediction shows no statistically significant improvement over random baselines under wet-lab label constraints.

VagueBench benchmark and training method show LLMs ask clarifying questions <11% of time; identifies proactive elicitation gap in 8B-frontier models.

NEIGHBOR LLM agent deployed in East Asian city for community services; aligns tool behavior with evidence and commitments.