Enneng Yang
Publications
ECHO: Entropy-Confidence Hybrid Optimization for Test-Time Reinforcement Learning
Test-time reinforcement learning generates multiple candidate answers via repeated rollouts and performs online updates using pseudo-labels constructed by majority voting. To reduce overhead and improve exploration, prior work introduces tree structured rollouts, which share reasoning prefixes and branch at key nodes to improve sampling efficiency. However, this paradigm still faces two challenges: (1) high entropy branching can trigger rollout collapse, where the branching budget concentrates on a few trajectories with consecutive high-entropy segments, rapidly reducing the number of effective branches; (2) early pseudo-labels are noisy and biased, which can induce self-reinforcing overfitting, causing the policy to sharpen prematurely and suppress exploration. To address these issues, we propose Entropy Confidence Hybrid Group Relative Policy Optimization (ECHO). During rollout, ECHO jointly leverages local entropy and group level confidence to adaptively control branch width, and further introduces online confidence-based pruning to terminate persistently low confidence branches, avoiding high entropy traps and mitigating collapse. During policy updates, ECHO employs confidence adaptive clipping and an entropy confidence hybrid advantage shaping approach to enhance training robustness and mitigate early stage bias. Experiments demonstrate that ECHO achieves consistent gains on multiple mathematical and visual reasoning benchmarks, and generalizes more effectively under a limited rollout budget.
Mitigating Safety Tax via Distribution-Grounded Refinement in Large Reasoning Models
Safety alignment incurs safety tax that perturbs a large reasoning model's (LRM) general reasoning ability. Existing datasets used for safety alignment for an LRM are usually constructed by distilling safety reasoning traces and answers from an external LRM or human labeler. However, such reasoning traces and answers exhibit a distributional gap with the target LRM that needs alignment, and we conjecture such distributional gap is the culprit leading to significant degradation of reasoning ability of the target LRM. Driven by this hypothesis, we propose a safety alignment dataset construction method, dubbed DGR. DGR transforms and refines an existing out-of-distributional safety reasoning dataset to be aligned with the target's LLM inner distribution. Experimental results demonstrate that i) DGR effectively mitigates the safety tax while maintaining safety performance across all baselines, i.e., achieving \textbf{+30.2\%} on DirectRefusal and \textbf{+21.2\%} on R1-ACT improvement in average reasoning accuracy compared to Vanilla SFT; ii) the degree of reasoning degradation correlates with the extent of distribution shift, suggesting that bridging this gap is central to preserving capabilities. Furthermore, we find that safety alignment in LRMs may primarily function as a mechanism to activate latent knowledge, as a mere \textbf{10} samples are sufficient for activating effective refusal behaviors. These findings not only emphasize the importance of distributional consistency but also provide insights into the activation mechanism of safety in reasoning models.
MMGRid: Navigating Temporal-aware and Cross-domain Generative Recommendation via Model Merging
Model merging (MM) offers an efficient mechanism for integrating multiple specialized models without access to original training data or costly retraining. While MM has demonstrated success in domains like computer vision, its role in recommender systems (RSs) remains largely unexplored. Recently, Generative Recommendation (GR) has emerged as a new paradigm in RSs, characterized by rapidly growing model scales and substantial computational costs, making MM particularly appealing for cost-sensitive deployment scenarios. In this work, we present the first systematic study of MM in GR through a contextual lens. We focus on a fundamental yet underexplored challenge in real-world: how to merge generative recommenders specialized to different real-world contexts, arising from temporal evolving user behaviors and heterogeneous application domains. To this end, we propose a unified framework MMGRid, a structured contextual grid of GR checkpoints that organizes models trained under diverse contexts induced by temporal evolution and domain diversity. All checkpoints are derived from a shared base LLM but fine-tuned on context-specific data, forming a realistic and controlled model space for systematically analyzing MM across GR paradigms and merging algorithms. Our investigation reveals several key insights. First, training GR models from LLMs can introduce parameter conflicts during merging due to token distribution shifts and objective disparities; such conflicts can be alleviated by disentangling task-aware and context-specific parameter changes via base model replacement. Second, incremental training across contexts induces recency bias, which can be effectively balanced through weighted contextual merging. Notably, we observe that optimal merging weights correlate with context-dependent interaction characteristics, offering practical guidance for weight selection in real-world deployments.