arXiv subjects

Feng Chang

Publications and source records attributed to Feng Chang.

9 recordsLinked to original sources

arXiv2026

ClawGym II: Exploring Black-Box RL on Agent Harness

Agent harnesses have substantially improved performance on long-horizon tasks by coordinating agent interactions with the environment. However, reinforcement learning through complex harnesses remains largely unexplored, as scaling such training to long-horizon agent tasks introduces fundamental challenges. In this work, we present a unified black-box RL framework for stable and scalable optimization of general agents through complex harnesses. Concretely, we first build a sandbox-based execution infrastructure that isolates task environments and harnesses within temporary sandboxes for large-scale concurrent rollouts. We then decouple policy optimization from opaque harness execution and place a serving proxy at the model boundary to capture model calls. To reconstruct multi-turn trajectories and improve training efficiency, we organize the captured calls into prefix trees and further adapt both critic-based PPO and critic-free GRPO to optimize over the recovered tree structure. Meanwhile, we maintain training-inference consistency throughout the optimization process. Finally, we introduce mix-harness training, allowing a single model to be jointly optimized by heterogeneous harnesses. With Qwen3-30A3B, black-box RL improves Pass@1 on ClawGym-Bench by 9.98 and 14.81 points through OpenClaw and Claude Code, respectively, while remaining stable over 200-400 optimization steps. Moreover, the framework yields consistent gains on more challenging tasks such as JobBench and OfficeQA. Overall, our framework enables effective, stable, and scalable optimization of general agents through black-box harnesses, supporting unified training across heterogeneous execution systems.

cs.CL↗

arXiv2026

FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents

Training deep search agents requires verifiable questions whose answers remain unavailable until sufficient evidence has been acquired through search. Existing synthesis methods often increase apparent difficulty by enriching graph structures, but structural complexity alone does not guarantee realized search difficulty: the intended search process can collapse through a cheaper identifying route. We formalize this gap with a shortcut-aware difficulty framework and identify four actionable shortcut risks: evidence co-coverage, single-clue selectivity, exposed constants, and prior-knowledge binding. To diagnose their realized effects, we use trajectory signatures including solving cost, answer hit time, and prior-shortcut rate. Guided by this framework, we introduce FORT, a Framework of Shortcut-Resistant Training-Data Synthesis. FORT constructs shortcut-resistant training data by controlling shortcut risks across entity selection, evidence graph construction, question formulation, and adversarial refinement. Experiments show that FORT induces longer pre-answer search and fewer shortcut patterns than existing open-source deep search datasets. Using the resulting trajectories, we train FORT-Searcher with supervised fine-tuning (SFT) only, and it achieves the best overall performance among comparable-size open-source search agents on challenging deep search benchmarks. Relevant resources will be made available at https://github.com/RUCAIBox/FORT-Searcher.

cs.CL↗

arXiv2026

TMAS: Scaling Test-Time Compute via Multi-Agent Synergy

Test-time scaling has become an effective paradigm for improving the reasoning ability of large language models by allocating additional computation during inference. Recent structured approaches have further advanced this paradigm by organizing inference across multiple trajectories, refinement rounds, and verification-based feedback. However, existing structured test-time scaling methods either weakly coordinate parallel reasoning trajectories or rely on noisy historical information without explicitly deciding what should be retained and reused, limiting their ability to balance exploration and exploitation. In this work, we propose TMAS, a framework for scaling test-time compute via multi-agent synergy. TMAS organizes inference as a collaborative process among specialized agents, enabling structured information flow across agents, trajectories, and refinement iterations. To support effective cross-trajectory collaboration, TMAS introduces hierarchical memories: the experience bank reuses low-level reliable intermediate conclusions and local feedback, while the guideline bank records previously explored high-level strategies to steer subsequent rollouts away from redundant reasoning patterns. Furthermore, we design a hybrid reward reinforcement learning scheme tailored to TMAS, which jointly preserves basic reasoning capability, enhances experience utilization, and encourages exploration beyond previously attempted solution strategies. Extensive experiments on challenging reasoning benchmarks show that TMAS achieves stronger iterative scaling than existing test-time scaling baselines, with hybrid reward training further improving scaling effectiveness and stability across iterations. Code and data are available at https://github.com/IQuestLab/tmas.

cs.AI↗

arXiv2026

ClawGym: A Scalable Framework for Building Effective Claw Agents

Claw-style environments support multi-step workflows over local files, tools, and persistent workspace states. However, scalable development around these environments remains constrained by the absence of a systematic framework, especially one for synthesizing verifiable training data and integrating it with agent training and diagnostic evaluation. To address this challenge, we present ClawGym, a scalable framework that supports the full lifecycle of Claw-style personal agent development. Concretely, we construct ClawGym-SynData, a diverse dataset of 13.5K filtered tasks synthesized from persona-driven intents and skill-grounded operations, paired with realistic mock workspaces and hybrid verification mechanisms. We then train a family of capable Claw-style models, termed ClawGym-Agents, through supervised fine-tuning on black-box rollout trajectories, and further explore reinforcement learning via a lightweight pipeline that parallelizes rollouts across per-task sandboxes. To support reliable evaluation, we further construct ClawGym-Bench, a benchmark of 200 instances calibrated through automated filtering and human-LLM review. Relevant resources have been released at https://github.com/ClawGym.

cs.CL↗

arXiv2026

Evolving Jailbreaks: Automated Multi-Objective Long-Tail Attacks on Large Language Models

Large Language Models (LLMs) have been widely deployed, especially through free Web-based applications that expose them to diverse user-generated inputs, including those from long-tail distributions such as low-resource languages and encrypted private data. This open-ended exposure increases the risk of jailbreak attacks that undermine model safety alignment. While recent studies have shown that leveraging long-tail distributions can facilitate such jailbreaks, existing approaches largely rely on handcrafted rules, limiting the systematic evaluation of these security and privacy vulnerabilities. In this work, we present EvoJail, an automated framework for discovering long-tail distribution attacks via multi-objective evolutionary search. EvoJail formulates long-tail attack prompt generation as a multi-objective optimization problem that jointly maximizes attack effectiveness and minimizes output perplexity, and introduces a semantic-algorithmic solution representation to capture both high-level semantic intent and low-level structural transformations of encryption-decryption logic. Building upon this representation, EvoJail integrates LLM-assisted operators into a multi-objective evolutionary framework, enabling adaptive and semantically informed mutation and crossover for efficiently exploring a highly structured and open-ended search space. Extensive experiments demonstrate that EvoJail consistently discovers diverse and effective long-tail jailbreak strategies, achieving competitive performance with existing methods in both individual and ensemble level.

cs.CR↗

arXiv2023

SafeLS: Toward Building a Lockstep NOEL-V Core

Safety-critical systems such as those in automotive, avionics and space, require appropriate safety measures to avoid silent data corruption upon random hardware errors such as those caused by radiation and other types of electromagnetic interference. Those safety measures must be able to prevent faults from causing the so-called common cause failures (CCFs), which occur when a fault produces identical errors in redundant elements so that comparison fails to detect the errors and a failure arises. The usual solution to avoid CCFs in CPU cores is using lockstep cores, so that two cores execute the same flow of instructions, but with some time staggering so that their state is never identical and faults can only lead to different errors, which are then detectable by means of comparison. This paper extends Gaisler's RISC-V NOEL-V core with lockstep; and presents future prospects for its use and distribution.

cs.AR↗

arXiv2023

Integrating features from lymph node stations for metastatic lymph node detection

Metastasis on lymph nodes (LNs), the most common way of spread for primary tumor cells, is a sign of increased mortality. However, metastatic LNs are time-consuming and challenging to detect even for professional radiologists due to their small sizes, high sparsity, and ambiguity in appearance. It is desired to leverage recent development in deep learning to automatically detect metastatic LNs. Besides a two-stage detection network, we here introduce an additional branch to leverage information about LN stations, an important reference for radiologists during metastatic LN diagnosis, as supplementary information for metastatic LN detection. The branch targets to solve a closely related task on the LN station level, i.e., classifying whether an LN station contains metastatic LN or not, so as to learn representations for LN stations. Considering that a metastatic LN station is expected to significantly affect the nearby ones, a GCN-based structure is adopted by the branch to model the relationship among different LN stations. At the classification stage of metastatic LN detection, the above learned LN station features, as well as the features reflecting the distance between the LN candidate and the LN stations, are integrated with the LN features. We validate our method on a dataset containing 114 intravenous contrast-enhanced Computed Tomography (CT) images of oral squamous cell carcinoma (OSCC) patients and show that it outperforms several state-of-the-art methods on the mFROC, maxF1, and AUC scores,respectively.

eess.IV↗

arXiv2022

End-to-End QoS for the Open Source Safety-Relevant RISC-V SELENE Platform

This paper presents the end-to-end QoS approach to provide performance guarantees followed in the SELENE platform, a high-performance RISC-V based heterogeneous SoC for safety-related real-time systems. Our QoS approach includes smart interconnect solutions for buses and NoCs, along with multicore interference-aware statistics units to, cooperatively, achieve end-to-end QoS.

cs.AR↗

arXiv2022

SafeSoftDR: A Library to Enable Software-based Diverse Redundancy for Safety-Critical Tasks

Applications with safety requirements have become ubiquitous nowadays and can be found in edge devices of all kinds. However, microcontrollers in those devices, despite offering moderate performance by implementing multicores and cache hierarchies, may fail to offer adequate support to implement some safety measures needed for the highest integrity levels, such as lockstepped execution to avoid so-called common cause failures (i.e., a fault affecting redundant components causing the same error in all of them). To respond to this limitation, an approach based on a software monitor enforcing some sort of software-based lockstepped execution across cores has been proposed recently, providing a proof of concept. This paper presents SafeSoftDR, a library providing a standard interface to deploy software-based lockstepped execution across non-natively lockstepped cores relieving end-users from having to manage the burden to create redundant processes, copying input/output data, and performing result comparison. Our library has been tested on x86-based Linux and is currently being integrated on top of an open-source RISC-V platform targeting safety-related applications, hence offering a convenient environment for safety-critical applications.

cs.AR↗