RLVR trains language models to reason by rewarding correct final answers. A widely used algorithm, group-relative policy optimization (GRPO), gives every token of a solution the same credit, so dense process rewards have been proposed to credit intermediate steps, for example with a Monte-Carlo reward that scores each...
Vikash Singh, Wei-Cong Chen, Debargha Ganguly et al.· 0 citations
Jev and its open replicas choose among the options a caller supplies with a single forward pass, fast enough to decide inside an agent loop. We ask what such a model cannot decide, and find a simple limit: Jev cannot make a decision that an LLM makes only by writing out intermediate steps. On a benchmark that varies on...
Wang Yang, Xin-Peng Li, Yan-Yan Zhang et al.· 0 citations
Post-training raises the 4B model's per-attempt accuracy and four-sample all-correct reliability and reduces average SMT calls from 4.10 to 2.52 and tokens per attempt from 18.97k to 12.72k.
Debargha Ganguly, Wei-Cong Chen, Mohsen Hariri et al.· 4 citations
While Vision-Language-Action (VLA) models enable flexible action generation, their generalization across diverse environmental elements, including manipulated objects, destinations, and backgrounds, is limited by the lack of diversity in robotic training data. Trained end-to-end on such data, VLAs tend to exploit visua...
Yan-Yan Zhang, Di-Sheng Liu, Xin-Peng Li et al.· arXiv.org· 0 citations
The ablations analyse why Foil works and turn the findings into design guidance for looped MoE: the returns of looping and of widening the expert layers amplify each other, routing confidence tracks healthy expert use better than load balance, and a sparse looped MoE should therefore use more experts per layer and more...
It is shown that a lightweight LLM-only DPO update on tiny single-object-pair synthetic data mitigates the bias, lifting four-way robust accuracy by up to 100 points on synthetic data, and by 68.1 points on broader evaluation datasets WhatsUp, SpatialMQA-Direct, and VSR.