Selected work

Representative Papers

Action-Sketcher: From Reasoning to Action via Visual Sketches for Long-Horizon Robotic Manipulation

Jan 04, 2026arXiv.org

This work addresses the challenges of spatial reference ambiguity, task decomposition difficulty, and opaque decision-making that arise when using natural language instructions for long-horizon robotic manipulation. To overcome these issues, the authors propose employing editable visual sketches as an explicit intermediate representation, establishing a closed-loop “perceive–reason–sketch–act” workflow that precisely maps linguistic intent onto scene geometry. Through a multi-stage curriculum training framework, the approach integrates modality alignment, language-to-sketch consistency constraints, and sketch-guided reinforcement imitation learning to enable adaptive policy learning and real-time action prediction. Evaluated in both simulated and real-world complex environments, the method significantly improves task success rates and dynamic robustness while supporting human-in-the-loop correction and yielding interpretable, intervenable decision processes.

2 citationsRead paper

SimpleDeepSearcher: Deep Information Seeking via Web-Powered Reasoning Trajectory Synthesis

May 22, 2025

Existing RAG systems face three key challenges in multi-step deep search: low-quality training data, distributional shift between simulated environments and real-world deployment, and high operational costs. This paper proposes a lightweight and efficient framework featuring the first data synthesis method grounded in authentic Web interactions, coupled with a multi-criteria trajectory filtering strategy to circumvent the high sample complexity and distribution mismatch inherent in reinforcement learning. Using only 871 carefully curated trajectories for supervised fine-tuning (SFT), our approach significantly outperforms RL-based baselines across five cross-domain deep search benchmarks. Our core contributions are: (1) establishing the first high-fidelity trajectory synthesis paradigm tailored to realistic web environments; (2) empirically validating the efficacy of SFT under extreme data scarcity; and (3) providing a practical, deployable pathway for deep search systems.

2 citationsRead paper

Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks

Jan 31, 2026

This study investigates whether Latent-CoT models, exemplified by CODI, genuinely perform implicit step-by-step reasoning. Employing interpretability techniques—including logit-lens decoding, linear probing, attention analysis, and activation patching—the work systematically examines how intermediate states are represented and propagated in polynomial iteration tasks. The analysis reveals, for the first time, that CODI constructs complete reasoning paths in short-hop tasks but shifts to relying on compressed shortcuts in long-hop settings, retaining only late-stage intermediate representations. This strategy proves highly fragile under distributional shifts or increased optimization difficulty, exposing a fundamental vulnerability in its reasoning process.

1 citationsRead paper

Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units

Jan 29, 2026

The causal origins of interpretable units—such as induction heads—in large language models remain poorly understood. This work proposes a scalable mechanistic data attribution framework that integrates influence functions with causal interventions to establish, for the first time, direct causal links between specific training examples and the emergence of such interpretable components. The study reveals that structured repetitive data plays a catalytic role in circuit formation and demonstrates a direct functional relationship between induction heads and in-context learning capabilities. By selectively intervening on a small set of high-influence training samples, the emergence of attention heads can be significantly modulated. Furthermore, the proposed data augmentation strategy consistently accelerates circuit convergence across different model scales.

1 citationsRead paper

Enhancing LLM Efficiency: Targeted Pruning for Prefill-Decode Disaggregation in Inference

Aug 29, 2025

To address the high computational and communication overhead caused by separating prefill and decoding stages in large language model (LLM) inference, this paper proposes a modular pruning framework tailored for Prefill-Decoding (PD) separation architectures. The method comprises two key components: (1) stage-aware block-level pruning and knowledge distillation, dynamically adapting sparsity patterns to the distinct computational characteristics of prefill and decoding; and (2) token-position-aware KV cache pruning, which selectively retains only the initial and final tokens’ KV states in critical layers, drastically reducing inter-stage data transfer volume. Experiments demonstrate consistent accuracy retention under both unified and separated deployment configurations. The approach achieves a 20.56% end-to-end inference speedup and reduces inter-stage bandwidth consumption by 4.95×. This work establishes a novel paradigm for efficient LLM inference at the edge and in distributed settings.

1 citationsRead paper

Recent publications

Latest Papers

UniTraffic-Agent: Unified Traffic Video Reasoning for AI City Challenge 2026 Track 3 with Two Out-of-Domain Evaluations

Aug 13, 2026

This work addresses the challenges of traffic video understanding—such as sparse events and highly variable viewpoints—that hinder accurate parsing of event evolution and critical interactions. We propose UniTraffic-Agent, the first unified agent framework for multi-task traffic video reasoning, integrating traffic anomaly detection, fisheye video event understanding, and pedestrian intent question answering. Built upon a multimodal large language model, our approach employs an “observe–reason–act–verify” workflow, enhanced with timestamp-aware visual evidence sampling and task-specific action adapters to enable joint reasoning and cross-domain generalization. Evaluated in the AI City Challenge 2026, UniTraffic-Agent achieves leading performance across all three tasks: ranking 16th in TAR (0.5780), 2nd in FETV (0.4884), and 4th in PSI-VQA (64.4161).

0 citationsRead paper

MOON: Multi-Objective OrthoNormalized Updates for Multitask Learning

Aug 12, 2026

This work addresses the inefficiency and task interference in multi-task learning that arise from neglecting the structural properties of model parameter matrices. It introduces matrix geometry into multi-objective optimization for the first time, proposing an orthogonally normalized gradient update mechanism grounded in matrix-valued steepest descent theory under the spectral–nuclear norm geometry. The method guarantees convergence to Pareto stationary points even in non-convex settings. Empirical evaluations across multiple benchmarks demonstrate substantial improvements in both optimization efficiency and multi-task performance. Theoretically, the approach achieves convergence rates of $\mathcal{O}(T^{-1/2})$ in the deterministic setting and $\mathcal{O}(T^{-1/4})$ under stochastic gradients.

0 citationsRead paper

StateFlow: Building, Evolving, and Accessing 3D World States for Previsualization

Aug 12, 2026

Existing generative previsualization approaches rely on one-shot image or video synthesis, which limits fine-grained, iterative control over scene structure, motion, camera parameters, and spatiotemporal dynamics. This work proposes a controllable and editable previsualization framework centered on an explicit, persistent 3D world state. It introduces a three-stage pipeline—construction, evolution, and access—featuring a novel prior-guided, conflict-aware dual-view initialization to build the 3D state. Structured state transitions enable user-intent-driven local edits and state reuse, eliminating the need for full-scene regeneration. Off-the-shelf video generation models are leveraged to enhance visual fidelity. The approach substantially improves both editing flexibility and visual quality, offering a powerful solution for high-quality, interactive video prototyping in film, gaming, and related creative domains.

0 citationsRead paper

VideoVIBE: A Video-Grounded Diagnostic Benchmark for One-Shot Interactive Website Generation

Aug 10, 2026

Existing methods struggle to perform fine-grained, interpretable quality diagnosis of interactive web pages generated in a single pass from natural language, often failing to pinpoint the root causes of errors. To address this limitation, this work introduces VideoVIBE—the first diagnostic benchmark based on human interaction videos—and proposes V2Lens, a training-free multi-agent system that enables fine-grained behavioral fidelity analysis through joint visual and code verification. Integrating video question answering, multimodal large language models, multi-agent reasoning, and source-code–behavior alignment, V2Lens significantly enhances diagnostic performance across 13 state-of-the-art Video MLLMs. Using Gemini-2.5-Flash as the baseline (64.54 accuracy), V2Lens improves accuracy to 71.72, thereby substantially increasing both behavioral consistency and interpretability of generated web pages.

0 citationsRead paper

SLIM-0.5B: Learning Action-Grounded Predictive Latents for Robot Manipulation

Aug 10, 2026

This work addresses key limitations of existing vision-language-action (VLA) models—namely parameter redundancy, high computational cost, and the inclusion of semantically irrelevant information—in robotic continuous control. The authors propose SLIM, a self-supervised latent interaction model with only 0.5B parameters, which uniquely integrates action reconstruction with future latent prediction. By leveraging masked trajectory prediction, SLIM learns predictive latent representations grounded in actions, jointly modeling observations, actions, and their dynamic relationships. Built upon a hybrid Transformer backbone and eschewing pixel-level modeling, SLIM enables efficient language-conditioned action generation via flow matching. Notably, it achieves competitive or superior performance compared to large-scale VLA and world models in both simulation and real-world settings—without requiring additional embodied pretraining—while operating at lower latency, reduced memory footprint, and significantly smaller model size.

0 citationsRead paper