"Representational Curvature Modulates Behavioral Uncertainty in LLMs" Most interpretability work studies what features live inside LLMs. But this paper studies something deeper, which is the geometry of the whole evolving representation. The key result is that representational
@askalphaxiv
-
Talk on Building and Evaluating Agentic LLM Systems for Science
By
–
If AI is so good right now, why aren't they scientists yet? Come join us for this AI4Science on alphaXiv talk: Building and Evaluating Agentic LLM Systems for Science. In this session, Prof. Dhruv Kumar (
@gargdhruv36
) and Dhruv Trehan (
@dhruvtrehan9
) will share recent -
DGPO Method Advances Token-Level Credit Assignment in RL
By
–
"DGPO: Distribution Guided Policy Optimization for Fine Grained Credit Assignment" RL for reasoning has a credit assignment problem. One reward basically gets spread across the whole chain-of-thought. This paper, DGPO, fixes that by turning policy deviation into a token-level
-
GLM-5V-Turbo: Native Foundation Model for Multimodal Agents
By
–
“GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents” Most multimodal agents today are still basically text agents with vision adapters, and a lot of reasoning failures are actually the model failing to see the interface, layout, chart, or object correctly.
-
New AI Alignment Training Research from Anthropic Fellows
By
–
a new paper from Anthropic Fellows Program! "Model Spec Midtraining: Improving How Alignment Training Generalizes" A lot of alignment training teaches models what to say, but not why those behaviors are right. So before normal alignment fine-tuning, this research trains the
-
MolmoAct2: AI Action Reasoning Models for Real-World Robotics
By
–
"MolmoAct2: Action Reasoning Models for Real-world Deployment" This paper from AI2 tackles open robot action model for real-world deployment. It combines a stronger embodied-reasoning backbone with a continuous action expert. A key component is the adaptive depth reasoning,
-
SkillSynth Builds Skill Graphs for Terminal Agent Training
By
–
"Toward Scalable Terminal Task Synthesis via Skill Graphs" Terminal-agent training needs more diverse workflows. This Tencent Hunyuan paper, SkillSynth, builds a graph of terminal scenarios and skills, samples paths through it, then turns those paths into executable tasks. It
-
Meituan Longcat Introduces Asynchronous RL for LLM Training
By
–
Asynchronous RL for LLM training. Meituan Longcat fixes the rollout bottleneck from long reasoning traces by keeping multiple policy versions alive at once. Long trajectories can now stay on their original policy, so training can keep moving without dropping samples or breaking
-
Speculative Decoding Accelerates RL Post-Training Rollouts
By
–
"Accelerating RL Post-Training Rollouts via System-Integrated Speculative Decoding" Speculative decoding for RL rollouts! This paper speeds up post-training without changing the target policy’s sampling distribution. So a draft model proposes multiple tokens, and the policy
-
GenLIP Trains ViT to Directly Predict Caption Tokens
By
–
"Let ViT Speak: Generative Language-Image Pre-training" Instead of contrastive image-text matching or adding a separate text decoder, this paper, GenLIP, trains a ViT to directly predict caption tokens from image patches with a standard next-token loss. A key component they use