Searcharxiv⌕ Search

arXiv subjects

Yukai Wang

Publications and source records attributed to Yukai Wang.

At least 19 recordsLinked to original sources

Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis

On-policy distillation (OPD) is a powerful paradigm for LLM post-training, yet its reliance on teacher logits restricts it to white-box teachers. In this paper, we study whether rubric-based supervision can provide a text-only feedback interface for black-box OPD, and what makes it effective. Our instantiation, Rubric-based On-Policy Distillation (ROPD), induces rubrics by contrasting teacher responses with the student's current rollouts and uses them to score these rollouts for on-policy optimization. We find that, under matched settings, text-only ROPD outperforms logit-based OPD methods on four competition-math benchmarks, reaching the best AIME24 accuracy of standard logit-based OPD with 9.6x fewer training samples. Our analysis yields three findings: rubric rewards track correctness far better than teacher log-likelihood (AUC 0.90 vs. 0.35); training with them repairs more failed criteria and regresses fewer passed ones than logit-based supervision; and contrasting teacher responses with student rollouts is central to rubric quality. These results position rubric-based OPD as a flexible, black-box-compatible alternative to logit-based OPD. Our implementation is available at https://github.com/Peregrine123/ROPD_official.

cs.LG↗

Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents

Recent studies of embodied agents based on frontier multimodal large language models (MLLMs) develop harness engineering around the MLLM, such as harnesses with post-trained policies and recursive self-improvement loops. Yet whether these additions exceed the native capability of the MLLM remains unexplored. We introduce SimplEvo to explore how far an MLLM can improve under direct control when it retains its execution experience in self-written text memories. It requires no weight updates, post-trained policies, or privileged information. The MLLM writes a coordinate-free task memory once from a single demonstration. Under a strict train-test protocol, it then creates a bounded working memory on training episodes only and refines it after each episode with lessons on procedure, completion, and recovery. Both memories stay fixed at test time. On EBench, SimplEvo lifts the success rate (SR) of GPT-6-Astra under direct control from 46.7% to 81.7%, and tops the RoboDojo leaderboard with 69.63% SR and a score of 75.32. Compared with memories that also store past visual frames, SimplEvo reaches a similar SR with text memory alone while using fewer tokens. With its memories fixed, SimplEvo also generalizes to unseen layouts and randomized scene configurations. Moreover, the working memory transfers across models: written by GPT-6-Astra, it can be reused as is by weaker MLLMs and improve their performance without any adaptation of their own.

cs.RO↗

Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy

Frontier multimodal large language models (MLLMs) can control robots without robot-specific training, but how their capabilities compare and interact with post-trained vision-language-action (VLA) and world action model (WAM) policies across manipulation demands remains to be thoroughly evaluated. In this report, we evaluate GPT-6-Astra on EBench as the core of an agent that controls a robot through an agent harness under the same visual and proprioceptive observation protocol as post-trained policies. We study two agent configurations: direct control, in which the agent generates waypoint commands directly, and harness with policy, in which it can also invoke a post-trained VLA or WAM policy for action execution. Under direct control, the agent achieves competitive overall performance relative to post-trained policies. It leads on short-horizon tasks and exhibits the least performance variation across generalization dimensions, whereas it underperforms post-trained policies on long-horizon tasks, tabletop manipulation, and high-precision contact-rich tasks. In the harness with policy, invoking a post-trained VLA or WAM combines the complementary strengths of both components, raising overall task success to the level of the strongest compared post-trained policies while substantially reducing token cost. However, performance remains below the per-task upper bound of the standalone components due to suboptimal policy invocation, fixed-instruction execution without subgoals, and grasp disruption when the agent resumes control. Overall, our study characterizes the capability boundary of a frontier MLLM agent and suggests directions for improving future agent harness design. Project page: https://internrobotics.github.io/EBench-Astra-Tech-Report/

cs.RO↗

TCMClinicalReason-Bench: Can Language Models Reason from Pathogenesis to Prescription over Real-World Clinical Cases?

Large language models (LLMs) can generate clinical narratives that are insufficiently grounded in patient-specific evidence. In traditional Chinese medicine (TCM), errors can propagate from etiology and pathogenesis through syndrome diagnosis and treatment principles to prescription generation. We developed TCMClinicalReason-Bench using 2,000 multicenter electronic health record cases to distinguish case-grounded responses from fluent but unsupported diagnostic and therapeutic conclusions. Five general-purpose and two TCM-specific LLMs were evaluated in zero-shot settings. An evidence-constrained rubric assessed seven diagnostic and therapeutic components and three cross-block relations, allowing case-supported alternatives. Qwen3.7-Plus with TCM retrieval served as the automated judge, alongside parallel blinded ratings by five senior TCM clinicians on a 600-case subset. Structural completeness was nearly saturated (99.3-100.0%), but normalized content scores ranged from 40.7% to 54.1%. The five general-purpose models averaged 50.0%, versus 41.3% for the two smaller TCM-specific models. Cross-block logic consistency ranged from 60.8% to 66.8% and correlated moderately with content across cases (Pearson's r = 0.515-0.656). Deficits were greatest in prescription generation, prescription analysis, and symptom-guided modification. In judge stress testing on 100 independent cases, perturbation detection rates across the three relations were 57%, 56%, and 31%, with contradictions detected more reliably than omissions. Separating component quality from cross-block consistency localizes failures missed by endpoint and completeness metrics and identifies where clinician oversight remains necessary.

cs.AI↗

Retrieval-Driven Memory Reconsolidation for Long-Term LLM Agents

Long-term memory is essential for LLM-based agents operating over extended interactions. Existing memory systems primarily update memory when new information arrives, treating retrieval as the endpoint of memory access rather than a driver of memory evolution. Consequently, retrieval feedback is rarely exploited to reorganize memory for future access continuously. Moreover, most existing approaches rely on predefined memory structures together with fixed retrieval pipelines, limiting the agent's ability to organize and evolve its own memory autonomously. Inspired by memory reconsolidation in cognitive neuroscience, we propose \textbf{REALM}, a \textbf{r}econsolidation-\textbf{e}volution \textbf{a}gentic \textbf{l}ong-term \textbf{m}emory framework. It models long-term memory as a continual lifecycle by autonomously organizing memories into a heterogeneous cognitive graph, retrieving evidence via adaptively composed graph-search atoms, and continually reconsolidating memories based on retrieval feedback. REALM achieves an average accuracy of 75.97\% on LoCoMo and 65.11\% on LongMemEval, outperforming the strongest baselines by 7.17 and 1.31 points respectively. Ablation studies confirm that memory reconsolidation consistently boosts performance, with further analyses revealing that it progressively reorganizes related memory units into more coherent local structures for collective evidence recall and utilization during reasoning. These results suggest that retrieval-driven memory reconsolidation provides an effective mechanism for continually evolving long-term memory in LLM agents.

cs.CL↗

EBench: Elemental Diagnosis of Generalist Mobile Manipulation Policies

We present EBench, a simulation benchmark that diagnoses generalist mobile manipulation policies beyond a single success-rate scalar. EBench comprises 26 diverse and challenging manipulation tasks annotated along 5 capability dimensions and 4 generalization dimensions. We evaluate state-of-the-art generalist manipulation models including $π_0$, $π_{0.5}$, XVLA, and InternVLA-A1, and reveal that the models exhibit strikingly different capability profiles: $π_{0.5}$ achieves the highest test success rate, the best train--test retention, and the strongest mobile manipulation performance; $π_0$ leads on dexterous fixed-base and high-precision tasks; XVLA and InternVLA-A1 exhibit complementary strengths across atomic skills and operating regimes. Beyond capability profiling, EBench analyzes the generalization ability from 4 representative perspectives, identifying the impact of different distribution shift factors. The results reveal strengths and weaknesses of models behind an overall score. We hope this benchmark offers a broad set of diagnostic signals to guide iteration on generalist manipulation models.

cs.RO↗

Many Representations as Sums of Three Prime Cubes

Let $F_k(n)$ be the number of unordered representations \[ n=p_1^k+p_2^k+\cdots +p_k^k \] by primes, with repetitions allowed. Erdős stated that $\limsup_{n\to\infty} F_3(n)=\infty$, but his proof appears not to have been published. A complete unconditional proof is given. The principal input is the classical Hecke equidistribution theorem for the CM Fermat cubic; the rest of the argument uses standard estimates for primes in arithmetic progressions and elementary counting. The argument used for \(k=3\) does not extend to the case \(k=4\). Nevertheless, by applying the Green--Tao--Ziegler theorem to the linear forms arising from an admissible binary quartic identity, \(\limsup_{n\to\infty}F_4(n)\ge2\) is shown.

math.NT↗

IOI: Decoupling Kinematics and Physics for Interactive World Models

Developing generalist embodied agents requires interactive environments providing visually realistic feedback and accurate action-conditioned dynamics. Interactive world models address this by simulating such complex dynamics. However, purely data-driven methods struggle to ensure precise control alignment and physically plausible visual feedback due to a lack of explicit structural constraints. To address this, we propose IOI, a hybrid interactive world model integrating analytical kinematic priors with learned physical dynamics. Unlike data-driven approaches prone to spatiotemporal drift, IOI introduces explicit kinematic guidance, computing forward kinematics from action sequences for accurate motion trajectories. These trajectories are rendered into synchronized front, side, and top orthographic projections, eliminating the need for extrinsic camera calibration. A Multi-view Kinematic Aggregation and Injection module fuses these geometric cues and injects them into the video generator, providing geometry-consistent guidance. Conditioning video generation on these deterministic trajectories establishes a synergy between the analytical simulator and the world model. Decoupling deterministic motion into the kinematic prior frees the generator to model stochastic physical interactions. Experiments on the RoboTwin benchmark validate IOI across kinematic fidelity, out-of-distribution (OOD) generalization, and policy evaluation. IOI achieves state-of-the-art simulation performance and robust zero-shot generalization to unseen OOD tasks. Furthermore, IOI serves as a reliable policy evaluator, yielding success rates closely aligning with ground-truth physics simulators. On real-world platforms, policies trained on IOI-synthesized data match those trained on teleoperation demonstrations, solidifying its practical value for embodied policy learning.

cs.RO↗

SMMBench: A Benchmark for Source-Distributed Multimodal Agent Memory

Existing benchmarks for multimodal memory reasoning largely evaluate systems within pre-assembled contexts, but under-evaluate whether agents can use evidence distributed across independently originated sources. We argue that source-distributed memory composition is an important and under-examined bottleneck in multimodal agent memory, especially when relevant evidence is fragmented across heterogeneous artifacts such as conversations, profiles, screenshots, tables, images, and documents. To address this gap, we introduce Source-distributed Multimodal Memory Benchmark(SMMBench), which measures whether agents can retrieve, align, and compose multimodal evidence scattered across multiple sources rather than reason within a single curated context. SMMBench evaluates four core capabilities: (1) cross-source multimodal reasoning; (2) conflict resolution; (3) preference reasoning; (4) memory-grounded action prediction. The benchmark contains 1877 samples grounded in 264 sources. Experiments on representative memory-style and retrieval-based baselines show that current systems still struggle on these capabilities, positioning source-distributed multimodal memory as an important and still under-evaluated challenge for multimodal agents. Our data are available at https://huggingface.co/datasets/HuacanChai/SMMBench.

cs.CL↗

AgentNoiseBench: Benchmarking Robustness of Tool-Using LLM Agents Under Noisy Condition

Recent advances in large language models have enabled LLM-based agents to achieve strong performance on a variety of benchmarks. However, their performance in real-world deployments often that observed on benchmark settings, especially in complex and imperfect environments. This discrepancy largely arises because prevailing training and evaluation paradigms are typically built on idealized assumptions, overlooking the inherent stochasticity and noise present in real-world interactions. To bridge this gap, we introduce AgentNoiseBench, a framework for systematically evaluating the robustness of agentic models under noisy environments. We first conduct an in-depth analysis of biases and uncertainties in real-world scenarios and categorize environmental noise into two primary types: user-noise and tool-noise. Building on this analysis, we develop an automated pipeline that injects controllable noise into existing agent-centric benchmarks while preserving task solvability. Leveraging this pipeline, we perform extensive evaluations across a wide range of models with diverse architectures and parameter scales. Our results reveal consistent performance variations under different noise conditions, highlighting the sensitivity of current agentic models to realistic environmental perturbations.

cs.AI↗

SafeMLRM: Demystifying Safety in Multi-modal Large Reasoning Models

The rapid advancement of multi-modal large reasoning models (MLRMs) -- enhanced versions of multimodal language models (MLLMs) equipped with reasoning capabilities -- has revolutionized diverse applications. However, their safety implications remain underexplored. While prior work has exposed critical vulnerabilities in unimodal reasoning models, MLRMs introduce distinct risks from cross-modal reasoning pathways. This work presents the first systematic safety analysis of MLRMs through large-scale empirical studies comparing MLRMs with their base MLLMs. Our experiments reveal three critical findings: (1) The Reasoning Tax: Acquiring reasoning capabilities catastrophically degrades inherited safety alignment. MLRMs exhibit 37.44% higher jailbreaking success rates than base MLLMs under adversarial attacks. (2) Safety Blind Spots: While safety degradation is pervasive, certain scenarios (e.g., Illegal Activity) suffer 25 times higher attack rates -- far exceeding the average 3.4 times increase, revealing scenario-specific vulnerabilities with alarming cross-model and datasets consistency. (3) Emergent Self-Correction: Despite tight reasoning-answer safety coupling, MLRMs demonstrate nascent self-correction -- 16.9% of jailbroken reasoning steps are overridden by safe answers, hinting at intrinsic safeguards. These findings underscore the urgency of scenario-aware safety auditing and mechanisms to amplify MLRMs' self-correction potential. To catalyze research, we open-source OpenSafeMLRM, the first toolkit for MLRM safety evaluation, providing unified interface for mainstream models, datasets, and jailbreaking methods. Our work calls for immediate efforts to harden reasoning-augmented AI, ensuring its transformative potential aligns with ethical safeguards.

cs.LG↗

Capsule Endoscopy Image Enhancement for Small Intestinal Villi Clarity

This paper presents, for the first time, an image enhancement methodology designed to enhance the clarity of small intestinal villi in Wireless Capsule Endoscopy (WCE) images. This method first separates the low-frequency and high-frequency components of small intestinal villi images using guided filtering. Subsequently, an adaptive light gain factor is generated based on the low-frequency component, and an adaptive gradient gain factor is derived from the convolution results of the Laplacian operator in different regions of small intestinal villi images. The obtained light gain factor and gradient gain factor are then combined to enhance the high-frequency components. Finally, the enhanced high-frequency component is fused with the original image to achieve adaptive sharpening of the edges of WCE small intestinal villi images. The experiments affirm that, compared to established WCE image enhancement methods, our approach not only accentuates the edge details of WCE small intestine villi images but also skillfully suppresses noise amplification, thereby preventing the occurrence of edge overshooting.

cs.CV↗

A Highlight Removal Method for Capsule Endoscopy Images

The images captured by Wireless Capsule Endoscopy (WCE) always exhibit specular reflections, and removing highlights while preserving the color and texture in the region remains a challenge. To address this issue, this paper proposes a highlight removal method for capsule endoscopy images. Firstly, the confidence and feature terms of the highlight region's edges are computed, where confidence is obtained by the ratio of known pixels in the RGB space's R channel to the B channel within a window centered on the highlight region's edge pixel, and feature terms are acquired by multiplying the gradient vector of the highlight region's edge pixel with the iso-intensity line. Subsequently, the confidence and feature terms are assigned different weights and summed to obtain the priority of all highlight region's edge pixels, and the pixel with the highest priority is identified. Then, the variance of the highlight region's edge pixels is used to adjust the size of the sample block window, and the best-matching block is searched in the known region based on the RGB color similarity and distance between the sample block and the window centered on the pixel with the highest priority. Finally, the pixels in the best-matching block are copied to the highest priority highlight removal region to achieve the goal of removing the highlight region. Experimental results demonstrate that the proposed method effectively removes highlights from WCE images, with a lower coefficient of variation in the highlight removal region compared to the Crinimisi algorithm and DeepGin method. Additionally, the color and texture in the highlight removal region are similar to those in the surrounding areas, and the texture is continuous.

cs.CV↗

Region Feature Descriptor Adapted to High Affine Transformations

To address the issue of feature descriptors being ineffective in representing grayscale feature information when images undergo high affine transformations, leading to a rapid decline in feature matching accuracy, this paper proposes a region feature descriptor based on simulating affine transformations using classification. The proposed method initially categorizes images with different affine degrees to simulate affine transformations and generate a new set of images. Subsequently, it calculates neighborhood information for feature points on this new image set. Finally, the descriptor is generated by combining the grayscale histogram of the maximum stable extremal region to which the feature point belongs and the normalized position relative to the grayscale centroid of the feature point's region. Experimental results, comparing feature matching metrics under affine transformation scenarios, demonstrate that the proposed descriptor exhibits higher precision and robustness compared to existing classical descriptors. Additionally, it shows robustness when integrated with other descriptors.

cs.CV↗

A Robust Error-Resistant View Selection Method for 3D Reconstruction

To address the issue of increased triangulation uncertainty caused by selecting views with small camera baselines in Structure from Motion (SFM) view selection, this paper proposes a robust error-resistant view selection method. The method utilizes a triangulation-based computation to obtain an error-resistant model, which is then used to construct an error-resistant matrix. The sorting results of each row in the error-resistant matrix determine the candidate view set for each view. By traversing the candidate view sets of all views and completing the missing views based on the error-resistant matrix, the integrity of 3D reconstruction is ensured. Experimental comparisons between this method and the exhaustive method with the highest accuracy in the COLMAP program are conducted in terms of average reprojection error and absolute trajectory error in the reconstruction results. The proposed method demonstrates an average reduction of 29.40% in reprojection error accuracy and 5.07% in absolute trajectory error on the TUM dataset and DTU dataset.

cs.CV↗

A Feature Matching Method Based on Multi-Level Refinement Strategy

Feature matching is a fundamental and crucial process in visual SLAM, and precision has always been a challenging issue in feature matching. In this paper, based on a multi-level fine matching strategy, we propose a new feature matching method called KTGP-ORB. This method utilizes the similarity of local appearance in the Hamming space generated by feature descriptors to establish initial correspondences. It combines the constraint of local image motion smoothness, uses the GMS algorithm to enhance the accuracy of initial matches, and finally employs the PROSAC algorithm to optimize matches, achieving precise matching based on global grayscale information in Euclidean space. Experimental results demonstrate that the KTGP-ORB method reduces the error by an average of 29.92% compared to the ORB algorithm in complex scenes with illumination variations and blur.

cs.CV↗

An Error-Matching Exclusion Method for Accelerating Visual SLAM

In Visual SLAM, achieving accurate feature matching consumes a significant amount of time, severely impacting the real-time performance of the system. This paper proposes an accelerated method for Visual SLAM by integrating GMS (Grid-based Motion Statistics) with RANSAC (Random Sample Consensus) for the removal of mismatched features. The approach first utilizes the GMS algorithm to estimate the quantity of matched pairs within the neighborhood and ranks the matches based on their confidence. Subsequently, the Random Sample Consensus (RANSAC) algorithm is employed to further eliminate mismatched features. To address the time-consuming issue of randomly selecting all matched pairs, this method transforms it into the problem of prioritizing sample selection from high-confidence matches. This enables the iterative solution of the optimal model. Experimental results demonstrate that the proposed method achieves a comparable accuracy to the original GMS-RANSAC while reducing the average runtime by 24.13% on the KITTI, TUM desk, and TUM doll datasets.

cs.CV↗

Hierarchical Forgery Classifier On Multi-modality Face Forgery Clues

Face forgery detection plays an important role in personal privacy and social security. With the development of adversarial generative models, high-quality forgery images become more and more indistinguishable from real to humans. Existing methods always regard as forgery detection task as the common binary or multi-label classification, and ignore exploring diverse multi-modality forgery image types, e.g. visible light spectrum and near-infrared scenarios. In this paper, we propose a novel Hierarchical Forgery Classifier for Multi-modality Face Forgery Detection (HFC-MFFD), which could effectively learn robust patches-based hybrid domain representation to enhance forgery authentication in multiple-modality scenarios. The local spatial hybrid domain feature module is designed to explore strong discriminative forgery clues both in the image and frequency domain in local distinct face regions. Furthermore, the specific hierarchical face forgery classifier is proposed to alleviate the class imbalance problem and further boost detection performance. Experimental results on representative multi-modality face forgery datasets demonstrate the superior performance of the proposed HFC-MFFD compared with state-of-the-art algorithms. The source code and models are publicly available at https://github.com/EdWhites/HFC-MFFD.

cs.CV↗