Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Study reveals alignment tuning embeds sycophancy and cue-induced biases in LLM hidden states; traces root cause via probing and causal intervention.
Activation steering enables fine-grained control over LLM reasoning trajectories by breaking self-loops via latent state manipulation.
VDAR-Router selects LLMs via verbalized query difficulty analysis for cost-performance-aware routing without embedding-only heuristics.
Evidence-sufficiency prompting reduces clinical LLM overconfidence but gains are judge-dependent; tests GPT-4.5, Claude Opus, Gemini, Grok on real data.
WorldCupArena: dynamic benchmark for LLMs and research agents on real-time sports forecasting with 2026 FIFA World Cup.
Self-distillation method improves rubric-based RL for LLMs by addressing train-inference mismatch in open-ended task optimization.
SelectInfer: neuron-level optimization framework for efficient LLM inference on edge devices via selective neuron loading.
Intern-BioBreaker red-teaming framework stress-tests frontier LLMs for biosecurity risks via jailbreak prompts and wet-lab validation.
The next time you apply for a job, AI may screen your résumé before any human sees it. But there’s good reason to question whether AI will judge you fairly. Researchers already know that LLMs pick up human biases from their training data. New research suggests that LLMs can…
Study evaluates open-weight LLMs for extracting structured CVE threat data from autonomous vehicle vulnerability text.
ActiveVision benchmark measures whether MLLMs perform active observation via 17 tasks requiring repeated visual perception.
Frontier LLMs fail at exact string copying due to positional encoding bias; 2D-RoPE organizes text as 2D grid to fix.
New benchmark evaluates frontier LLMs on real analytical knowledge work—synthesizing information, judgment under uncertainty, strategic thinking—beyond factual recall and coding.
In-place tokenizer expansion for LLMs: reallocate vocabulary post-training to reduce latency/energy for underrepresented languages.
Expert-in-the-loop annotation framework for depression diagnosis using LLMs, grounding labels in DSM-5-TR criteria for explainable mental health AI.
Benchmark evaluates six MLLMs on scientific visualization literacy using 49 standardized assessment items across 8 visualization techniques.
Deep Interaction: efficient human-in-the-loop mechanism for correcting reasoning errors in LLMs via targeted intervention rather than full regeneration.
PAT: RAG-based system for whole-document translation beyond sentence-level via corpus-informed context, targeting professional draft translation workflows.
AIMO Interpretability Challenge: competition to distinguish robust from spurious reasoning in frontier math LLMs via mechanistic analysis.
Training-free human-object interaction detection using multimodal LLMs for open-world compositional scenarios without dataset-specific supervision.
Method to certify LLM honesty under incentive pressure via counterfactual report mediators that resist manipulation while remaining responsive to evidence.
Placebo-controlled methodology (PoPE) to measure whether small code LLMs can actually use execution error feedback to repair code.
Pythia multi-agent system for autonomous clinical symptom extraction using open-weights LLMs without fine-tuning.
Meta-evaluation of LLM-generated rubrics for paper reproduction tasks; first systematic study on checklist-based evaluation reliability.
KLLMs pretraining paradigm suppresses parametric recall via entity anonymization, forcing evidence-grounded reasoning over memorized knowledge.
Survey of efficient inference techniques for diffusion-based LLMs, covering algorithms and systems beyond theoretical parallel generation claims.
LLMs fine-tuned to predict chemical reaction mechanisms step-by-step, reducing hallucinations versus name-reaction prediction approaches.
ESFP benchmark measures whether LLMs distinguish and coherently shift between neutral attribution and self-stance epistemic registers in contested claims.
Elenchos framework evaluates abductive reasoning in LLMs via formal-system mutation detection, exposing gap between pattern recognition and latent-hypothesis inference.
Cost-aware speculative decoding for Mixture-of-Experts LLMs optimizes expert activation patterns to reduce inference cost beyond token acceptance alone.