Yuheng Zhang, Mingyue Huo, Minghao Zhu, Mengxue Zhang, Nan Jiang
Reward models (RMs) are widely used as optimization targets in reinforcement learning from human feedback (RLHF). We show that leading RMs can assign higher scores to gibberish than to meaningful reference answers. To systematically discover such inputs, we introduce Token Mapping Perturbation Attack (TOMPA), a framework that uses a perturbed token mapping to extend an attack LLM's search beyond natural language. The mapping changes the linguistic meaning of the attack LLM's natural language responses, producing non-linguistic text for the target RM to score. We then use these scalar scores as rewards to train the attack LLM through reinforcement learning to discover non-linguistic adversarial inputs. Experiments against Llama and Qwen RMs demonstrate such attacks using two instantiations: identity token mapping across different tokenizers and random permutation with a shared tokenizer. Against Skywork-Reward-V2-Llama-3.1-8B, ranked first on RewardBench 2, the two mappings achieve mean rewards of +33.64 and +43.78, respectively, compared with +17.48 for GPT-5 reference answers. Their attack success rates are 97.6% and 100%, measured as the percentages of individual responses scoring above their corresponding GPT-5 references. These results show that strong benchmark performance does not ensure robustness to non-linguistic adversarial inputs, exposing a fundamental vulnerability in using RMs as optimization targets.
Yuheng Zhang, Wenlin Yao, Changlong Yu, Yao Liu, Qingyu Yin, Bing Yin, Hyokun Yun, Lihong Li
Group relative policy optimization (GRPO) trains large language models with verifiable rewards by normalizing rewards within response groups. When every response in a group receives the same reward, the advantages vanish and the group provides no policy gradient signal. We propose Adaptive Rollout and Response Reuse Policy Optimization (AR3PO) to obtain informative groups with fewer generations. Adaptive rollout samples responses to each prompt in stages and stops sampling once the model generates a correct response. If all sampled responses are incorrect, response reuse replaces one of them with a past correct response to the same prompt, when available, and recomputes the group advantages. Policy-recomputed reuse trains on the updated groups using probabilities recomputed under the rollout policy; advantage-only reuse uses historical responses only to compute advantages. Across four mathematical benchmarks with 7B and 8B models, AR3PO consistently improves over GRPO and achieves performance comparable to DAPO while requiring up to $4.2\times$ fewer generated responses. Experiments with a 32B model also show comparable performance to DAPO at a similar training stage with substantially fewer generated responses.
We study post-training for dynamically routed mixture-of-experts (MoE) models, focusing on integrating independently optimized specialists while keeping consolidation efficient. Our approach combines calibrated multi-teacher on-policy distillation (C-MOPD), which matches each selected teacher on a local Top-$K$ support with teacher-wise loss normalization and clipping, with an execution path that uses post-capacity bounded-width dispatch, final-route replay, and asynchronous rollout--training overlap. On Turing-20B-A2B, the selected model achieves a four-domain normalized integration score of 0.9988 and retains at least 90.16\% of the measured specialist gain in every domain. At matched $K=32$, the full calibration objective scores 0.9962, compared with 0.9724 without calibration and 0.9834 without clipping. At a matched 8K response cap, the model exceeds both shared SFT and Qwen3-8B on all four reasoning-enabled mathematics and code benchmarks (unweighted mean: 85.92, versus 80.33 for Qwen3-8B and 74.21 for SFT), and improves over SFT on ten of twelve broader evaluation metrics. In a PPU-based profile, the combined execution configuration reduces consolidation step time from 398.7 to 139.3 seconds, an observed 2.86$\times$ speedup over serial execution. These results demonstrate strong specialist integration alongside more efficient consolidation, while reflecting the complete training and execution framework rather than any single component.
Large language model (LLM) distillation can propagate teacher bias under covariate shift when target reward feedback is unavailable. We propose Iterative Calibration and Learning (IterCaL), which alternates source-reward teacher calibration through token-level branching with student learning on target prompts. We prove polynomial convergence in expected average KL to the oracle student, the reference-regularized target-reward maximizer within a potentially misspecified student class; a quantified source-to-target calibration discrepancy gives an explicit additive true-return term. Our separation shows that regularized direct teacher matching retains positive oracle-KL error at its population optimum for every shared reference penalty, even when the teacher is stronger than every student. On the same instance, IterCaL converges to the oracle using source rewards alone.
Yuheng Zha, Yilei Wang, Qiyue Gao, Junrong Chen, Yujia Wu, Zhengfeng Lai, Zhengzhong Liu, Eric P. Xing
Humans often solve spatial problems by mentally simulating visual transformations. In contrast, conventional vision-language models (VLMs) reason primarily through language. We investigate whether VLMs can solve spatial problems by reasoning with both text and generated visual states. To this end, we introduce WM-VLM, which equips a pretrained VLM with a lightweight world model branch for generating intermediate visual states. Our two-stage training first teaches the model to generate the next visual state and then to use that state for reasoning. We programmatically construct spatial reasoning tasks with verifiable intermediate visual states. These tasks allow us to evaluate how well the model generates visual states and how much it relies on them to answer the question. On 2D and 3D mental rotation tasks, WM-VLM consistently outperforms the supervised fine-tuned backbone, with gains of up to 39.25 percentage points. Ablations suggest that these gains depend on the generated visual states, as removing or corrupting them sharply reduces performance. Together, these results suggest that internal world models offer a promising path toward VLMs that reason in both language and visual space. Code is publicly available at: https://github.com/yuh-zha/WM-VLM
Mingyue Huo, Yiwen Shao Nuance Labs, Yuheng Zhang University of Illinois, Urbana Champaign
We present TagSpeech, a unified LLM-based framework that utilizes Temporal Anchor Grounding for joint multi-speaker ASR and diarization. The framework is built on two key designs: (1) decoupled semantic and speaker streams fine-tuned via Serialized Output Training (SOT) to learn turn-taking dynamics; and (2) an interleaved time anchor mechanism that not only supports fine-grained timestamp prediction but also acts as a synchronization signal between semantic understanding and speaker tracking. Compared to previous works that primarily focus on speaker-attributed ASR or implicit diarization, TagSpeech addresses the challenge of fine-grained speaker-content alignment and explicitly models who spoke what and when in an end-to-end manner. Experiments on AMI and AliMeeting benchmarks demonstrate that our method achieves consistent improvements in Diarization Error Rate (DER) over strong end-to-end baselines, including Qwen-Omni and Gemini, particularly in handling complex speech overlaps. Moreover, TagSpeech employs a parameter-efficient training paradigm in which the LLM backbone is frozen and only lightweight projectors are trained, resulting in strong performance with low computational cost
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表