The Archive

Non-linear intervention framework extends LLM mechanistic understanding beyond Linear Representation Hypothesis to implicitly encoded features.

Empirical study evaluates whether LLMs correctly interpret formal semantics of High-Level Message Sequence Charts.

MILM framework applies LLMs to irregular multimodal time series (e.g., EHR) with informative sampling strategy.

ORCE method decouples verbalized confidence from answer generation in LLMs to improve uncertainty calibration without degrading accuracy.

Q-DAPS estimates question difficulty for LLM QA evaluation using entropy of answer plausibility scores.

TokenHD: token-level hallucination detection pipeline for reasoning tasks with improved granularity over step-based methods.

Analysis of popularity bias in OLMo LLMs traced to pretraining exposure in Dolma corpus via entity-level statistics.

GAP framework fixes feature-space mismatch in multimodal LLM visual reasoning by aligning latent token generation with input embedding norms.

δ-mem adds compact online associative memory to LLM backbones via delta-rule updates, enabling efficient long-context reuse in agentic systems.

Evaluates LLM ability to generate Entity-Relationship diagrams from natural language; identifies systematic limitations in database modeling tasks.

Qwen-Scope: open-source suite of sparse autoencoders for mechanistic interpretability across 7 Qwen models.

EvoNav uses LLMs to evolve reward functions for robot navigation via reinforcement learning, automating design of navigation policies.

DISCA steers LLM cultural preferences via sociodemographic disagreement signals without fine-tuning or white-box access.

Psych-201 dataset reveals post-training reduces LLM behavioral alignment with humans, with divergence widening in newer model generations.

Method to compress SFT-induced LLM behaviors into sparse, causally necessary subnetworks for selective inference-time control.

UniSD unifies self-distillation design choices for LLM adaptation without external teachers, systematizing complementary techniques.

Systematic review of jailbreak attack and defense methods for LLMs with critique of narrow evaluation metrics like attack success rate.

Single-pass hallucination detection method for LLMs using attention head KL-divergence without sampling, validated across multiple model families.

Comparative study of LoRA and QLoRA fine-tuning on Bashkir, a low-resource Turkic language, using models from DistilGPT2 to Qwen2.5-7B.

SaFE-Scale framework reveals clinical LLM safety and accuracy follow divergent scaling laws; introduces RadSaFE-2 benchmark.

A new study examines how large language models perform in a variety of medical contexts, including real emergency room cases — where at least one model seemed to be more accurate than human doctors.

MolJSON: novel molecular representation format systematically compared against SMILES/IUPAC for LLM chemistry reasoning tasks.

ML-Bench&Guard: policy-grounded multilingual safety benchmark (14 languages) aligning LLMs with region-specific regulations and cultural context.

SC-Taxo generates hierarchical scientific taxonomies using LLMs with semantic consistency constraints across hierarchy levels.

LLM-based combinatorial optimization for Design Structure Matrix modularization; engineering application without novel AI contribution.

DPN-LE framework identifies minimally necessary neurons for LLM personality representation to reduce editing overhead and degradation.

TIDE framework enables cross-architecture knowledge distillation for diffusion LLMs, reducing parameter requirements while maintaining parallel decoding capabilities.

Benchmark evaluating 72 LLMs on 270 harmful instructions for robotic health attendant safety; mean violation rate 54.4% across models.

Experimental study demonstrating LLMs can be manipulated to prioritize fringe scientific material and generate misleading fluent responses contradicting scientific consensus.

First systematic study of uncertainty estimation in audio-aware LLMs; benchmarks five methods addressing hallucination and confidence calibration.