Selected work

Representative Papers

MathArena: Evaluating LLMs on Uncontaminated Math Competitions

May 29, 2025

Public benchmark datasets (e.g., AIME 2024) suffer from widespread data leakage, confounding LLM mathematical reasoning evaluation with memorization effects. Method: We introduce the first contamination-free benchmark for mathematical reasoning, built on real-time released contest problems—149 unseen questions from five major competitions (AIME, SMT, USAMO, etc.)—governed by a strict decontamination protocol aligned with official contest release windows. Contribution/Results: We propose a novel multi-granularity scoring scheme jointly evaluating answer correctness and proof rigor, enabling the first standardized, systematic assessment of formal proof generation. Experiments reveal that while state-of-the-art models excel on uncontaminated problem-solving tasks (e.g., SMT 2025), their scores drop below 25% on USAMO 2025 proof-generation tasks—unambiguously exposing deficiencies in formal deductive reasoning. This stark performance gap validates the benchmark’s efficacy in disentangling genuine reasoning capability from dataset memorization.

12 citationsRead paper

Incentivizing Honesty among Competitors in Collaborative Learning and Optimization

May 25, 2023Neural Information Processing Systems

This paper addresses the problem of strategic participants in collaborative learning who submit misleading model updates to gain competitive advantage, thereby degrading others’ performance. Unlike prior work assuming malicious adversaries, we model participants as self-interested yet non-malicious rational agents and propose the first incentive-compatible mechanism for federated learning. Leveraging game-theoretic analysis and mechanism design, we rigorously prove that truthful updates constitute a Nash equilibrium—both in single-round mean estimation and multi-round strongly convex SGD. Theoretically, the mechanism guarantees learning performance approaching that of full cooperation. Empirically, we validate its effectiveness on non-convex federated benchmarks (e.g., FEMNIST): it substantially mitigates strategic manipulation, achieves convergence rates comparable to fully cooperative training, and seamlessly integrates with standard FedAvg without architectural modification.

3 citationsRead paper

Deterministic Negative-Weight Shortest Paths in Nearly Linear Time via Path Covers

Nov 11, 2025

For the single-source shortest paths (SSSP) problem with negative edge weights and negative cycle detection in directed graphs, all prior near-linear-time algorithms rely on low-diameter decompositions and are randomized. This paper presents the first deterministic near-linear-time algorithm, achieving a time complexity of $ ilde{O}(m log(nW))$, which matches the optimal bound for deterministic SSSP in such graphs. The key innovation is the introduction of *path covering*—a novel structural primitive—that enables the first complete derandomization of low-diameter-decomposition-based approaches. Leveraging the integrality of edge weights and an efficient path-covering construction, our method avoids traditional random sampling entirely. This resolves a long-standing open problem in deterministic graph algorithm design and provides a scalable new tool for optimization on directed graphs.

1 citationsRead paper

Parameter Exploration for RLVR via Variational Learning

Aug 10, 2026

This work addresses a key limitation in current reinforcement learning approaches for large language models, which predominantly rely on action-space exploration—such as temperature scaling—and struggle to effectively reorder tokens, often leading to training divergence or stagnation. To overcome this, the paper introduces Perturbed Parameter Policy Optimization (3PO), the first systematic framework leveraging parameter-space exploration. Built upon a variational formulation of the policy posterior, 3PO generates diverse trajectories through parameter perturbations and enhances exploration efficiency via a reward-based grouping mechanism. Evaluated on OLMo-3-1025-7B and Qwen2.5-Math-7B across mathematical reasoning and code generation tasks, 3PO consistently outperforms standard GRPO, yielding substantial gains in downstream performance with negligible computational overhead while significantly reducing zero-advantage groups and erroneous outputs.

0 citationsRead paper

Seeing Is Not Deciding: Can Multimodal LLMs Act as Effective CEOs?

Aug 06, 2026

Existing benchmarks inadequately assess the ability of multimodal large language models to integrate visual business evidence in executive decision-making. To address this gap, this work introduces C-SUITEBENCH, a multimodal benchmark that systematically evaluates nine state-of-the-art models across 50 real-world business scenarios under both textual and multimodal conditions, focusing on CEO-level strategic decisions. Through controlled experiments, ablation studies, and visual channel analyses, the study uncovers a “multimodal integration paradox”: while visual information substantially enhances risk prediction and argumentation capabilities, it consistently impairs constrained decisions such as resource allocation, primarily due to signal crowding effects. These findings reveal that perception and action constitute distinct bottlenecks in multimodal agents, highlighting a critical divergence between interpretive and operational intelligence in complex business contexts.

0 citationsRead paper

Recent publications

Latest Papers

Parameter Exploration for RLVR via Variational Learning

Aug 10, 2026

This work addresses a key limitation in current reinforcement learning approaches for large language models, which predominantly rely on action-space exploration—such as temperature scaling—and struggle to effectively reorder tokens, often leading to training divergence or stagnation. To overcome this, the paper introduces Perturbed Parameter Policy Optimization (3PO), the first systematic framework leveraging parameter-space exploration. Built upon a variational formulation of the policy posterior, 3PO generates diverse trajectories through parameter perturbations and enhances exploration efficiency via a reward-based grouping mechanism. Evaluated on OLMo-3-1025-7B and Qwen2.5-Math-7B across mathematical reasoning and code generation tasks, 3PO consistently outperforms standard GRPO, yielding substantial gains in downstream performance with negligible computational overhead while significantly reducing zero-advantage groups and erroneous outputs.

0 citationsRead paper

Seeing Is Not Deciding: Can Multimodal LLMs Act as Effective CEOs?

Aug 06, 2026

Existing benchmarks inadequately assess the ability of multimodal large language models to integrate visual business evidence in executive decision-making. To address this gap, this work introduces C-SUITEBENCH, a multimodal benchmark that systematically evaluates nine state-of-the-art models across 50 real-world business scenarios under both textual and multimodal conditions, focusing on CEO-level strategic decisions. Through controlled experiments, ablation studies, and visual channel analyses, the study uncovers a “multimodal integration paradox”: while visual information substantially enhances risk prediction and argumentation capabilities, it consistently impairs constrained decisions such as resource allocation, primarily due to signal crowding effects. These findings reveal that perception and action constitute distinct bottlenecks in multimodal agents, highlighting a critical divergence between interpretive and operational intelligence in complex business contexts.

0 citationsRead paper

iARCS: Iterative Agentic RL for Controllable 3D Scene Generation

Aug 06, 2026

Existing 3D scene generation methods struggle to reliably satisfy task-critical functional constraints such as navigability and reachability, limiting the practical utility of synthetic data. This work proposes an iterative agent-based reinforcement learning framework that first enhances physical plausibility and layout quality through pretraining with generic rewards, then leverages a large language model (LLM) to generate executable, task-specific reward programs. These LLM-generated rewards are integrated into a feedback-driven reinforcement learning loop for iterative refinement. By uniquely combining LLM-synthesized reward functions with iterative reinforcement learning, the approach significantly improves adherence to functional constraints while preserving scene diversity, thereby enhancing downstream task performance.

0 citationsRead paper

Towards Compact Unified Multimodal Tracking: Synergizing Knowledge Distillation with Structural Pruning

Aug 02, 2026

This work addresses the challenges of high computational overhead and redundant prediction heads in unified multimodal object tracking models, which hinder efficient deployment on edge devices. To this end, the authors propose an efficient compression method that combines structured pruning of the decoder with a novel dual-alignment knowledge distillation framework. This framework decouples knowledge transfer into spatial representation alignment—emphasizing target localization—and semantic distribution alignment—preserving discriminative dark knowledge. Evaluated across RGB, thermal, and depth modalities, the approach is systematically assessed against 17 distillation strategies. The proposed method achieves state-of-the-art performance on five benchmarks, attaining a 91.5% MPR on RGBT234 while running at 54 FPS on a single RTX 4090 GPU—offering a 5× speedup over the baseline with improved accuracy.

0 citationsRead paper

DVPSFormer: Efficient Online Depth-aware Video Panoptic Segmentation for Autonomous Driving

Jul 28, 2026

This work addresses the high computational cost and reliance on offline tracking in existing depth-aware video panoptic segmentation methods, which hinder real-time deployment in autonomous driving. To overcome these limitations, the authors propose a unified online framework that integrates semantic, instance, and metric depth estimation within a single forward pass. The approach employs an Explicit Scene Discretization (ESD) mechanism to distinguish foreground from background and introduces a Discrete-to-Continuous (D2C) depth decoder to recover accurate metric depth. Furthermore, an Online Majority Voting (OMV) mechanism is incorporated to enhance temporal consistency and classification robustness. This method achieves state-of-the-art performance on Cityscapes-DVPS and SemKITTI-DVPS while significantly reducing latency, marking the first efficient solution for online 4D scene understanding suitable for real-time autonomous perception.

0 citationsRead paper