Turbo Harness: Instance-Adaptive Harness Optimization
Tunyu Zhang*, Hao Wang*, Kai Xu, Dimitris N. Metaxas
Paper | Code

• Turbo Harness tailors an agent’s harness to each task instance, turning past optimization experience into better agent execution.
• A lightweight, RL-trained harness editor uses a reusable playbook to patch a globally optimized harness, with just one editor call per task.
• Improves performance across seven benchmarks spanning interactive agents, software engineering, and long-horizon terminal tasks, while often reducing execution steps and costs.

Sequential Probabilistic Uncertainty Estimation for Parallel Multi-Agent Reasoning Systems
Tunyu Zhang, Zihao Zhao, Yusong Zhao, Haizhou Shi, Zhuohang Li, Haoxian Chen, Hao Wang, Dimitris N. Metaxas
Paper | Code | Data

• SAUCE estimates the reliability of multi-agent reasoning by tracking how consensus evolves throughout an interaction.
• Combines agent agreement and generation uncertainty in a training-free sequential estimator, using existing inference traces without extra sampling.
• Improves error detection, selective prediction, and calibration across five model backbones, five benchmarks, and two multi-agent protocols.

Few-Step Diffusion Language Models via Trajectory Self-Distillation
Tunyu Zhang*, Xinxi Zhang*, Ligong Han, Haizhou Shi, Xiaoxiao He, Zhuowei Li, Hao Wang, Kai Xu, Akash Srivastava, Chengzhi Mao, Hao Wang, Vladimir Pavlovic, Dimitris N. Metaxas
Paper | Code | Slides

• T3D enables high-quality text generation in fewer diffusion steps by distilling a full-step teacher’s generative trajectory into a few-step student.
• Trajectory supervision reduces token factorization error, while Direct Discriminative Optimization (DDO) further strengthens reasoning through a mode-seeking objective.
• Substantially closes the quality gap between few-step and full-step decoding on reasoning and code generation while preserving full-step performance.

TokUR: Token-Level Uncertainty Estimation for Large Language Model Reasoning
Tunyu Zhang*, Haizhou Shi*, Yibin Wang, Hengyi Wang, Xiaoxiao He, Zhuowei Li, Haoxian Chen, Ligong Han, Kai Xu, Huan Zhang, Dimitris Metaxas, Hao Wang
Paper | Code

• We propose TokUR, a framework for token-level uncertainty estimation tailored for LLM reasoning.
• TokUR introduces a low-rank stochastic perturbation mechanism to approximate predictive distributions efficiently.
• The framework enables more reliable multi-step reasoning, and provides uncertainty-aware signals for downstream tasks.

When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning
Zihao Zhao, Tunyu Zhang, Haizhou Shi, Yusong Zhao, Xinxi Zhang, Hao Wang
Paper | Code

• Explains when multi-agent debate outperforms majority voting: agents must propose a correct answer, and the final decision must recognize it even when it comes from a minority.
• Introduces Latent Verification Debate (LVD) to model how verification evidence helps recover correct proposals that voting misses.
• Selects agents with complementary answer coverage, improving reasoning accuracy across two model backbones under matched inference budgets.

Multimodal needle in a haystack: Benchmarking long-context capability of multimodal large language models
Hengyi Wang, Haizhou Shi, Shiwei Tan, Weiyi Qin, Wenyuan Wang, Tunyu Zhang, Akshay Nambi, Tanuja Ganu, Hao Wang

Paper | Code

• MMNeedle provides a systematic evaluation framework for long-context multimodal understanding.
• It enables controlled benchmarking of retrieval and reasoning over large visual contexts, and reveals robustness challenges in current multimodal LLMs.