Searcharxiv⌕ Search

arXiv subjects

Xiangyang Zhou

Publications and source records attributed to Xiangyang Zhou.

11 recordsLinked to original sources

WordInferBench: Can Agents Infer How to Edit from Word Document Context?

Document editing is among the most common knowledge-work tasks, yet existing benchmarks for evaluating autonomous agents on Word document editing provide fully specified attribute-modification instructions. They do not evaluate agents' ability to infer the appropriate attributes from document context, a capability essential in real-world editing where users expect consistency without spelling out every attribute. We introduce WordInferBench, a benchmark comprising 280 base document-editing instances, each paired with three instruction conditions: descriptive (the instruction explicitly lists every attribute change), referential (the instruction points to another object as an exemplar), and inferential (the agent must examine document context to infer what attributes to apply). To evaluate contextual attribute inference, each base instance is constructed so that the target formatting is recoverable from an in-document exemplar. The instances are derived from real Word documents collected from Common Crawl and cover 13 object types and 81 attributes. We evaluate six recent coding-agent systems and find that the best-performing evaluated system, Claude Code with Opus 4.6, achieves a 62.9% task-level pass rate on descriptive variants but only 40.7% on inferential variants, indicating that current systems complete tasks less reliably when the required formatting changes must be determined from document context.

cs.AI↗

WordEval: Evaluating Word-Native Operation Fidelity in Document Editing

Microsoft Word is a central environment for office automation, and the faithful execution of its editing operations is an important capability for document agents. Unlike plain-text editing, editing in Word modifies a structured object model, where the requested effect is realized in object properties. Existing benchmarks, however, evaluate office workflows or desktop automation, treating word-processor documents as one task domain among many and verifying outcomes with keyword matching, per-task property checks, or screenshot comparison. We introduce WordEval, an operation-fidelity benchmark that instead centers evaluation on Word editing operations, testing whether each requested edit is realized as intended changes to Word objects and properties. It contains 642 cases across 15 Word feature areas, ranging from text and lists to charts, shape geometry, and SmartArt. 546 atomic cases cover 231 commands, 96 compositional cases combine multiple edits in a single document, and a Word-native oracle checks task-specific object properties and preservation conditions. Cases are constructed from editing-unit specifications, yielding controlled input documents, reference documents, and aligned oracle checks. Each compositional case is linked to corresponding atomic cases, supporting fine-grained diagnosis of editing failures. We evaluate eight agent systems and find that many struggle on the Atomic track, with performance varying across feature areas. Moreover, every system scores lower on Compositional than on Atomic, with gaps that vary substantially by system. Exact atomic-intent links and recorded failure diagnoses provide complementary views of this variation.

cs.CV↗

M$^\star$: Every Task Deserves Its Own Memory Harness

Large language model agents rely on a memory harness to write, organize, retrieve, and use past experience. A harness that works for one task often fails on another, because conversations, embodied planning, and expert reasoning require different storage and retrieval behavior. To address this limitation, we introduce Mstar, a method that automatically discovers task-optimized memory programs through reflective code evolution. Specifically, Mstar operationalizes a memory harness as an executable Python memory program that defines the Schema, Logic, and Instruction, and optimizes these components jointly. We evaluate Mstar on four tasks covering conversation, embodied planning, and specialized reasoning. Our results demonstrate that Mstar improves performance over existing baselines across all evaluated tasks. Furthermore, the evolved memory programs exhibit structurally distinct processing mechanisms for each evaluated domain. These findings suggest that every task benefits from its own memory harness, and that memory-program search provides a concrete way to discover it.

cs.PL↗

CART: Closed-Loop Adaptive Red Teaming for Large Language Models

Automated red teaming often replays a fixed set of prompts, which measures known risks but cannot learn from failures found during testing. We present CART (Closed-Loop Adaptive Red Teaming), a framework that uses each result to guide what it tests next. CART begins with broad risk coverage, follows weaknesses that emerge, keeps new probes diverse, and records the evidence and source of every finding. It separates the Challenger that creates tests, the Target being tested, which may be a text-only model or a bounded tool-using agent, and the Judge that evaluates the results, allowing these roles to be studied independently. Across three evaluation families (Frontier, JAH, and Agentic), CART discovers more failures and higher average risk than static seed replay for every Target with an available baseline. The gains extend to tool-mediated agent tests, suggesting that contextual adaptation can reveal weaknesses that direct prompt replay does not exercise. These results describe what the test policies discover, not how often failures occur in real deployments. We also find that Challenger-Judge choices affect the evidence uncovered, highlighting the need for role separation and independent review. Overall, CART turns red teaming from a one-time checklist into a continuous, adaptive, and auditable search for model and agent weaknesses.

cs.AI↗

Evolving Agents in the Dark: Retrospective Harness Optimization via Self-Preference

AI agents rely on a harness of skills, tools, and workflows to solve complex problems. Continually improving this harness is essential for adapting to new tasks. However, existing optimization methods typically require ground-truth validation sets, yet such labeled data is difficult to acquire in practical deployment settings. To address this problem, we introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes the agent harness using only past trajectories. Specifically, RHO selects a diverse coreset of challenging tasks from past trajectories and re-solves them in parallel. The agent analyzes these rollouts using self-validation and self-consistency, then generates candidate harness updates and selects the most effective one by its own pairwise self-preference. We evaluate RHO across three diverse domains, spanning software engineering, technical work, and knowledge work. Notably, a single optimization round improves the pass rate on SWE-Bench Pro from 59% to 78% without any external grading. Furthermore, our analysis demonstrates that RHO effectively targets prior failure modes. As a result, the optimized harness alters the agent's behavior patterns and sustains higher accuracy during long-horizon sessions.

cs.AI↗

Mind the Gap: Can Frontier LLMs Pass a Standardized Office Proficiency Exam?

The deployment of Large Language Model (LLM) agents for computer automation is accelerating, yet their ability to navigate complex, professional-grade productivity software is largely untested. We argue that Office automation is an ideal environment for benchmarking document-automation capability, as it requires long-horizon planning and reasoning, precise parameter configuration, and multi-application integration. To quantify this capability, we introduce an evaluation based on China's National Computer Rank Examination (NCRE), featuring 200 comprehensive practical-operation tasks across Word, Excel, and PowerPoint. Each task is scored on a 100-point rubric scale using 7,118 machine-gradable criteria, and Score Rate (SR) denotes the mean percentage of rubric points earned across these tasks. We benchmark 7 frontier LLMs and observe stark limitations: single-turn models score a maximum of 36.6%. A stronger agentic system with execution feedback, iterative repair, and broader Office automation access reaches 68.8%, but remains below the 95.5% community-reference score used as a scoring sanity check. Ultimately, our experiments demonstrate that despite recent advancements in code generation, achieving reliable fine-grained Office document automation remains a significant challenge for current code-generating LLM and agent systems.

cs.AI↗

The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions

Large Language Models' safety-aligned behaviors, such as refusing harmful queries, can be represented by linear directions in activation space. Previous research modeled safety behavior with a single direction, limiting mechanistic understanding to an isolated safety feature. In this work, we discover that safety-aligned behavior is jointly controlled by multi-dimensional directions. Namely, we study the vector space of representation shifts during safety fine-tuning on Llama 3 8B for refusing jailbreaks. By studying orthogonal directions in the space, we first find that a dominant direction governs the model's refusal behavior, while multiple smaller directions represent distinct and interpretable features like hypothetical narrative and role-playing. We then measure how different directions promote or suppress the dominant direction, showing the important role of secondary directions in shaping the model's refusal representation. Finally, we demonstrate that removing certain trigger tokens in harmful queries can mitigate these directions to bypass the learned safety capability, providing new insights on understanding safety alignment vulnerability from a multi-dimensional perspective. Code and artifacts are available at https://github.com/BMPixel/safety-residual-space.

cs.CL↗

Distilling Knowledge from Pre-trained Language Models via Text Smoothing

This paper studies compressing pre-trained language models, like BERT (Devlin et al.,2019), via teacher-student knowledge distillation. Previous works usually force the student model to strictly mimic the smoothed labels predicted by the teacher BERT. As an alternative, we propose a new method for BERT distillation, i.e., asking the teacher to generate smoothed word ids, rather than labels, for teaching the student model in knowledge distillation. We call this kind of methodTextSmoothing. Practically, we use the softmax prediction of the Masked Language Model(MLM) in BERT to generate word distributions for given texts and smooth those input texts using that predicted soft word ids. We assume that both the smoothed labels and the smoothed texts can implicitly augment the input corpus, while text smoothing is intuitively more efficient since it can generate more instances in one neural network forward step.Experimental results on GLUE and SQuAD demonstrate that our solution can achieve competitive results compared with existing BERT distillation methods.

cs.CL↗

How to Evaluate the Next System: Automatic Dialogue Evaluation from the Perspective of Continual Learning

Automatic dialogue evaluation plays a crucial role in open-domain dialogue research. Previous works train neural networks with limited annotation for conducting automatic dialogue evaluation, which would naturally affect the evaluation fairness as dialogue systems close to the scope of training corpus would have more preference than the other ones. In this paper, we study alleviating this problem from the perspective of continual learning: given an existing neural dialogue evaluator and the next system to be evaluated, we fine-tune the learned neural evaluator by selectively forgetting/updating its parameters, to jointly fit dialogue systems have been and will be evaluated. Our motivation is to seek for a lifelong and low-cost automatic evaluation for dialogue systems, rather than to reconstruct the evaluator over and over again. Experimental results show that our continual evaluator achieves comparable performance with reconstructing new evaluators, while requires significantly lower resources.

cs.CL↗

Proactive Human-Machine Conversation with Explicit Conversation Goals

Though great progress has been made for human-machine conversation, current dialogue system is still in its infancy: it usually converses passively and utters words more as a matter of response, rather than on its own initiatives. In this paper, we take a radical step towards building a human-like conversational agent: endowing it with the ability of proactively leading the conversation (introducing a new topic or maintaining the current topic). To facilitate the development of such conversation systems, we create a new dataset named DuConv where one acts as a conversation leader and the other acts as the follower. The leader is provided with a knowledge graph and asked to sequentially change the discussion topics, following the given conversation goal, and meanwhile keep the dialogue as natural and engaging as possible. DuConv enables a very challenging task as the model needs to both understand dialogue and plan over the given knowledge graph. We establish baseline results on this dataset (about 270K utterances and 30k dialogues) using several state-of-the-art models. Experimental results show that dialogue models that plan over the knowledge graph can make full use of related knowledge to generate more diverse multi-turn conversations. The baseline systems along with the dataset are publicly available

cs.CL↗

Power-Law Graph Cuts

Algorithms based on spectral graph cut objectives such as normalized cuts, ratio cuts and ratio association have become popular in recent years because they are widely applicable and simple to implement via standard eigenvector computations. Despite strong performance for a number of clustering tasks, spectral graph cut algorithms still suffer from several limitations: first, they require the number of clusters to be known in advance, but this information is often unknown a priori; second, they tend to produce clusters with uniform sizes. In some cases, the true clusters exhibit a known size distribution; in image segmentation, for instance, human-segmented images tend to yield segment sizes that follow a power-law distribution. In this paper, we propose a general framework of power-law graph cut algorithms that produce clusters whose sizes are power-law distributed, and also does not fix the number of clusters upfront. To achieve our goals, we treat the Pitman-Yor exchangeable partition probability function (EPPF) as a regularizer to graph cut objectives. Because the resulting objectives cannot be solved by relaxing via eigenvectors, we derive a simple iterative algorithm to locally optimize the objectives. Moreover, we show that our proposed algorithm can be viewed as performing MAP inference on a particular Pitman-Yor mixture model. Our experiments on various data sets show the effectiveness of our algorithms.

cs.CV↗