Weitao Feng

Haoqin Tu, Weitao Feng, Hardy Chen et al.

Process-supervised reward models serve as a fine-grained function that provides detailed step-wise feedback to model responses, facilitating effective selection of reasoning trajectories for complex tasks. Despite its advantages, evaluation on PRMs remains less explored, especially in the multimodal domain. To address this gap, this paper first benchmarks current vision large language models (VLLMs) as two types of reward models: output reward models (ORMs) and process reward models (PRMs) on multiple vision-language benchmarks, which reveal that neither ORM nor PRM consistently outperforms across all tasks, and superior VLLMs do not necessarily yield better rewarding performance. To further advance evaluation, we introduce ViLBench, a vision-language benchmark designed to require intensive process reward signals. Notably, OpenAI's GPT-4o with Chain-of-Thought (CoT) achieves only 27.3% accuracy, indicating the benchmark's challenge for current VLLMs. Lastly, we preliminarily showcase a promising pathway towards bridging the gap between general VLLMs and reward models -- by collecting 73.6K vision-language process reward data using an enhanced tree-search algorithm, our 3B model is able to achieve an average improvement of 3.3% over standard CoT and up to 2.5% compared to its untrained counterpart on ViLBench by selecting OpenAI o1's generations. We release the implementations at https://ucsc-vlaa.github.io/ViLBench with our code, model, and data.

Weitao Feng, Lixu Wang, Tianyi Wei et al.

As large language models (LLMs) continue to grow in capability, so do the risks of harmful misuse through fine-tuning. While most prior studies assume that attackers rely on supervised fine-tuning (SFT) for such misuse, we systematically demonstrate that reinforcement learning (RL) enables adversaries to more effectively break safety alignment and facilitate advanced harmful task assistance, under matched computational budgets. To counter this emerging threat, we propose TokenBuncher, the first effective defense specifically targeting RL-based harmful fine-tuning. TokenBuncher suppresses the foundation on which RL relies: model response uncertainty. By constraining uncertainty, RL-based fine-tuning can no longer exploit distinct reward signals to drive the model toward harmful behaviors. We realize this defense through entropy-as-reward RL and a Token Noiser mechanism designed to prevent the escalation of expert-domain harmful capabilities. Extensive experiments across multiple models and RL algorithms show that TokenBuncher robustly mitigates harmful RL fine-tuning while preserving benign task utility and finetunability. Our results highlight that RL-based harmful fine-tuning poses a greater systemic risk than SFT, and that TokenBuncher provides an effective and general defense.

Weitao Feng, Zhihao Hu, Baopu Li et al.

Multi-Object Tracking (MOT) is a popular topic in computer vision. However, identity issue, i.e., an object is wrongly associated with another object of a different identity, still remains to be a challenging problem. To address it, switchers, i.e., confusing targets thatmay cause identity issues, should be focused. Based on this motivation,this paper proposes a novel switcher-aware framework for multi-object tracking, which consists of Spatial Conflict Graph model (SCG) and Switcher-Aware Association (SAA). The SCG eliminates spatial switch-ers within one frame by building a conflict graph and working out the optimal subgraph. The SAA utilizes additional information from potential temporal switcher across frames, enabling more accurate data association. Besides, we propose a new MOT evaluation measure, Still Another IDF score (SAIDF), aiming to focus more on identity issues.This new measure may overcome some problems of the previous measures and provide a better insight for identity issues in MOT. Finally,the proposed framework is tested under both the traditional measures and the new measure we proposed. Extensive experiments show that ourmethod achieves competitive results on all measure.

Weitao Feng, Zhihao Hu, Wei Wu et al.

In this paper, we propose a unified Multi-Object Tracking (MOT) framework learning to make full use of long term and short term cues for handling complex cases in MOT scenes. Besides, for better association, we propose switcher-aware classification (SAC), which takes the potential identity-switch causer (switcher) into consideration. Specifically, the proposed framework includes a Single Object Tracking (SOT) sub-net to capture short term cues, a re-identification (ReID) sub-net to extract long term cues and a switcher-aware classifier to make matching decisions using extracted features from the main target and the switcher. Short term cues help to find false negatives, while long term cues avoid critical mistakes when occlusion happens, and the SAC learns to combine multiple cues in an effective way and improves robustness. The method is evaluated on the challenging MOT benchmarks and achieves the state-of-the-art results.