Amortising Bayesian Experimental Design for Sequential Information Gathering in LLMs
ASIG: fine-tuning LLMs for sequential information gathering via amortized Bayesian Experimental Design and information-gain rewards.
Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
ASIG: fine-tuning LLMs for sequential information gathering via amortized Bayesian Experimental Design and information-gain rewards.
Spectral shape-based metrics using Heavy-Tailed Self-Regularization theory to fingerprint and compare LLMs across architecture/scale/training differences.
Risk-controlled real-time safety monitor for LLM outputs at deployment time, thresholding verifier signals to flag unsafe generations with mathematical guarantees.
DramaSR-532K benchmark for speaker recognition in TV dramas with 532K annotated dialogue lines; applies reasoning LLMs to improve multimodal attribution.
Scaling laws investigation suggests LLM social simulation fidelity may not improve with compute scaling; proposes dedicated research attention.
Philosophical critique of LLMs as cultural measurement tools, arguing models constitute rather than neutrally measure cultural phenomena.
Comparative evaluation of frontier LLMs (GPT, Claude Opus, Gemini, GLM) for automated Linux/bash exam grading using cognitive taxonomy.
NeuFS uses neuron-level signals to identify high-value samples for few-shot LLM adaptation, reducing annotation costs in domain specialization.
SPG-Layout generates physically plausible indoor scenes from text in non-orthogonal environments using LLMs and statistical priors.
RFM-AGOP efficiently extracts multi-dimensional refusal subspaces in LLMs for safety steering, handling long reasoning traces computationally.
WattGPU predicts GPU power draw and inter-token latency for LLM inference across unseen hardware using only public specs, enabling deployment optimization.
DecompRL teaches LLMs modular code generation via RL to solve problems with near-zero base policy success rate, scaling accuracy beyond sampling limits.
DRIFTLENS measures reasoning drift in personalized LLMs; memory injection reshapes inference trajectories on open-ended questions without ground truth.
Survey identifies limitations of LLM-as-a-Judge evaluation in multilingual and low-resource language settings with inadequate validation.
Rubric-based clinical reasoning benchmark shows frontier LLMs achieve ~32% on hard expert-authored medical cases.
AutoMem treats memory management as trainable skill, enabling LLMs to autonomously decide file-system operations and knowledge organization.
Cartridge distillation detects stealth biases in LLMs by exposing preferential signals in logit distributions invisible to text-based analysis.
ZO-Act enables zeroth-order fine-tuning of LLMs via activation-informed low-rank subspace optimization without backpropagation.
University RAG-based multimodal chatbot prototype for institutional Q&A, combining LLMs with semantic retrieval for domain-specific policy queries.
Evan Feinberg and Sergey Edunov discuss diffusion models for drug discovery at Genesis Molecular AI, including PEARL's OpenBind performance and protein co-folding advances.
Let’s start with a game. Open up your chatbot of choice—Claude, ChatGPT, Gemini—and type “Give me a random number between 1 and 10.” You’re going to get 7. Almost always. Now type “Another” and you’ll get 3 or 4. Type “Another” again and you’ll get 8 or 9. That won’t work every…
Zero-shot benchmark evaluates Claude, GPT-5.4, and Gemini on fine-grained 13-class emotion classification.
Berkeley BAIR lab celebrates 2026 Ph.D. cohort with work spanning robotics, LLMs, vision, safety, and AI for science.
RL with metacognitive feedback trains LLMs to accurately express uncertainty and improve performance through self-monitoring.
Systematic evaluation reveals data referencing errors (1.7B-20B models misread tables) as overlooked reliability issue beyond final-answer accuracy.
CoMet decomposes uncertainty sources in multimodal LLMs via context and multiplicity factorization for better calibration in open-ended tasks.
Surrogate Fidelity evaluates when mechanistic interpretability findings from open models transfer to closed model predictions, attributions, and representations.
NCP-ToM benchmark evaluates LLM agents' ability to induce belief states through planning/action beyond conversation.
Step-aware RL for medical multimodal LLMs addresses cascading reasoning failures in clinical VQA via fine-grained credit assignment.
Evo-PI framework evolves principle-guided supervision signals for multimodal LLMs to improve medical reasoning generalization.