9 papers · 1 filter
Smaller Models, Better Rejects: Preference Distillation Scaling
Rui Cai, Wenhui Zhu, Xiwen Chen +15
Preference distillation typically treats a teacher response as preferred and the student's own response as rejected. This assumes that self-generated failures are the most informat…
TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning
Yuanda Xu, Zhengze Zhou, Hejian Sang +6
Agentic reinforcement learning requires assigning credit to environment-facing actions such as searches, clicks, edits, navigation commands, and object interactions. Standard GRPO…
TREK: Distill to Explore, Reinforce to Refine
Yuanda Xu, Zhengze Zhou, Kayhan Behdin +10
Group Relative Policy Optimization (GRPO) is effective when the current policy already samples useful reasoning trajectories, but it stalls on hard prompts whose correct solution m…
TIP: Token Importance in On-Policy Distillation
Yuanda Xu, Hejian Sang, Zhengze Zhou +3
On-policy knowledge distillation (OPD) trains a student on its own rollouts under token-level supervision from a teacher. Not all token positions matter equally, but existing views…
Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training
Yuanda Xu, Hejian Sang, Zhengze Zhou +3
In settings where labeled verifiable training data is the binding constraint, each checked example should be allocated to the model and reward density where it is most informative.…
Robust Batch-Level Query Routing for Large Language Models under Cost and Capacity Constraints
Jelena Markovic-Voronov, Kayhan Behdin, Yuanda Xu +3
We study the problem of routing queries to large language models (LLMs) under cost, GPU resources, and concurrency constraints. Prior per-query routing methods often fail to contro…
CRISP: Compressed Reasoning via Iterative Self-Policy Distillation
Hejian Sang, Yuanda Xu, Zhengze Zhou +3
Reasoning models often generate far more tokens than a task requires, which raises inference cost and can compound errors. We introduce CRISP (Compressed Reasoning via Iterative Se…
Overconfident Errors Need Stronger Correction: Asymmetric Confidence Penalties for Reinforcement Learning
Yuanda Xu, Hejian Sang, Zhengze Zhou +2
Reinforcement Learning with Verifiable Rewards (RLVR) has become the leading paradigm for enhancing reasoning in Large Language Models (LLMs). However, standard RLVR algorithms suf…
Not all tokens are needed(NAT): token efficient reinforcement learning
Hejian Sang, Yuanda Xu, Zhengze Zhou +2
Reinforcement learning (RL) has become a key driver of progress in large language models, but scaling RL to long chain-of-thought (CoT) trajectories is increasingly constrained by…