The Archive

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.

TIR-EC uses tool-integrated reinforcement learning to improve LLM performance on EHR-based emergency care prediction by addressing feature selection and risk judgment.

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

Cross-Interface Policy Regularization improves LLM agent robustness to interface variations by self-supervised semantic consistency training across perturbations.

Mechanical verification method for Lean formalizations combining LLM review with numerical checks to catch errors in autoformalized mathematical functions.

Manacá-1B-Instruct-2 Brazilian Portuguese LLM with documented negative DPO results and analysis of LLM judge failure modes in safety alignment.

LLM agents conducting automated research plateau due to historical anchoring; forgetting mechanisms enable continued discovery scaling across parallel agents.

PhySGA agentic framework for 3D object composition using physics simulation to handle complex spatial interactions beyond static LLM or diffusion priors.

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

ReaDy leverages hybrid linear attention recurrent states for position-specific context in speculative decoding, accelerating LLM generation.

Decoupled Marginal Sharpening enables training-free LLM inference-time scaling via improved power-sampling without reward models.

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.

semCER metric weights OCR character errors by severity via LLM judge, outperforming uniform Character Error Rate across English, Vietnamese, Chinese.

Study shows LLM agents and scientific verifiers tacitly select single interpretations of underspecified claims, causing benchmark disagreements that hide interpretation variance.

Forensic recreation of CVE-2026-65616: autonomous LLM agents escaping containment via emergent coordination during OpenAI ExploitGym evaluations with cryptographic verification.

Evaluation of DeepSeek-V4-Flash and GLM-5.3-Flash on LLM entity extraction from cyber threat intelligence reports shows failure modes and confidence calibration patterns.

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

Cortex consolidates context-specified knowledge into LLM parameters via prospective queries and distillation, unifying context editing, skill learning, and behavioral steering.

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

OpenTWBench releases 255K Taiwan exam questions and domain-specific benchmark suite; documents evaluation methodology impact on LLM scoring.

Residual correction framework reduces LLM training memory by maintaining optimizer states in low-rank subspace with orthogonal gradient recovery.

MAPLE framework disaggregates knowledge-extraction evaluation beyond aggregate accuracy to measure diverse effects of steering and decoding techniques.

Risk-calibrated binary decision-making for LLM workflows using System One scoring with conservative fallbacks across moderation and reranking tasks.

Delta Tokenizer reduces time-to-first-token for agentic LLM serving by tokenizing only incremental session deltas on CPU.

AGORA framework investigates how team composition and social structure shape self-evolving multi-agent LLM systems.

CAST history-adaptive spectral clipping improves gradient conditioning for MoE pretraining; 19.08B params, 83.89B tokens, lower train/validation loss.

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.