The Archive

Technique enabling Multimodal LLMs to construct 3D scene representations from multi-view images before answering visual reasoning questions, improving cross-viewpoint coherence.

Quantization technique for hybrid LLMs with linear attention (Gated DeltaNet, Kimi Delta Attention) addressing outlier rows/columns and rounding error accumulation in recurrent states.

Small trainable advisor guides frozen frontier LLMs via natural-language guidance, learning from task rewards and interaction feedback.

Study shows audio encoder pruning in Speech-LLMs increases demographic disparities in WER across groups on Fair-Speech and Common Voice.

Power-sharpened sampling balances exploration-exploitation at inference time to enhance reasoning in LLMs without RL post-training.

RICE enables training-free dense retrieval by prompting LLMs with in-context examples to produce effective query-document representations.

Method for quantifying LLM reasoning uncertainty using divergent token analysis beyond raw token probabilities for confidence calibration.

Study of gender bias across 10 LLMs from 9 vendors shows heterogenous bias patterns using stereotype attribution and moral judgment tasks.

State says LLMs threaten civilization as "the greatest public nuisance ever created."

Failure-Transparent Agents (FTA) benchmark isolates post-failure reporting errors in tool-using LLMs via controlled failure states and evidence auditing.

AI Night-Scientist uses reinforcement learning to teach LLMs to balance structured reasoning with creative divergence for scientific discovery.

MedEVM benchmark reveals LLMs misalign diagnostic decisions from evidence quality in dynamic medical tasks across 1,050 cases.

Simon Willison summarizes major LLM developments in 2026, including Claude Opus 4.5 and GPT-5.1 releases and industry trends.

FIRE stabilizes LLM self-distillation via Fisher-informed recalibration of feedback on correct/incorrect outputs.

Study reveals probabilistic incoherence in LLMs (Jev, English Laya) when decomposing classification decisions across hierarchical category levels.

Routing drift in merged MoE LLMs does not necessarily indicate failure; proposes analysis toolkit for counterfactual routing assessment.

Clinical LLMs exhibit faithfulness gaps: high accuracy without correct reasoning. Introduces ESI, CFS, PSS metrics for explanation fidelity.

OpenTumorBoard: 611-case benchmark from 12,534 minutes of real tumor board recordings evaluating LLMs on specialist turn prediction and board simulation.

Clinical LLM evaluation reveals six models fail counterfactual validity and demographic robustness tests: 22% of CFV attempts and significant demographic bias detected.

Belief Self-Distillation extracts and manipulates user models from LLMs via compact learnable representations.

DeepEdu-v1: Vietnamese-native open-weights tutoring agent addressing data sovereignty and curriculum alignment for education in developing regions.

EAServe: disaggregated serving system optimizing Encode-Prefill-Decode pipeline for multimodal LLMs with resource-aware request flow regulation.

First systematic evaluation of cultural awareness in LLMs for Haitian Creole across specificity, bias, diversity, and variation dimensions.

Kaggle Game Arena evaluates LLMs through competitive games (Chess, Poker, Werewolf) preventing benchmark saturation.

ViSTA adapter integrates irregular clinical time-series into vision-language models for structured patient data prediction and reasoning.

Intent2Tc uses LLMs to translate business-level QoS intents into validated Linux traffic-control policies via closed-loop framework.

Highlight-Then-Summarize trains LLMs to compress sparse evidence before reasoning on long contexts; H2S-Dataset covers 11 benchmark families.

Prism's larger goal is open-weight AI that runs on devices and makes better use of the computing power they already have.

Study across 13 MLLMs shows standard similarity metrics misattribute visual-language alignment; noise-injection exposes shallow integration.

LLMs generate MRS meaning representations better than they parse them using ERG, reconstructing 10K benchmark tasks bidirectionally.