Jun Nie, Yonggang Zhang, Qianshu Cai, Yiu-ming Cheung, Xinmei Tian, Bo Han
An LLM agent can acquire persistent improvements by evolving its harness---the executable program governing model calls, tool use, verification, and recovery---without updating model weights. When the same harness serves users with different workloads, isolated evolution limits the sharing of useful discoveries, while centralized adaptation can introduce conflicting task-specific changes. We introduce EvolveNet, a framework for collaborative harness evolution that connects local specialization with shared capability accumulation. Clients evolve the same broadcast harness on their local workloads in parallel, and a server composes their adaptations into an updated shared program. Redistribution makes accepted discoveries available to all clients, allowing subsequent local evolution to build on improvements accumulated across the collaboration. EvolveNet supports this process through evidence-guided, scope-typed program aggregation and behavioral validation. Our theoretical analysis provides sufficient conditions for retaining and combining complementary improvements despite program interactions. Experiments across five settings show improved shared-harness performance, with ablation and behavioral analyses supporting the contribution of composition beyond specialist selection alone.
Jun Nie, Yonggang Zhang, Jun Song, Qianshu Cai, Dahai Yu, Yike Guo, Xinmei Tian, Bo Han
The behavior of an LLM agent is determined not only by the underlying model, but also by its harness: the executable program that constructs context, invokes tools, verifies intermediate results, and recovers from failures. Existing approaches typically optimize harnesses on training or development data and freeze the resulting programs at test time, limiting adaptation to shifts in task distributions, failure modes, and tool interactions. We ask whether the harness can instead be optimized during evaluation using unlabeled execution traces. We introduce Test-Time Harness Evolution (TTHE), which treats the executable harness as the state of test-time adaptation. On each unlabeled test batch, TTHE iteratively evolves a population of candidate harnesses through agentic proposers that reason over execution traces and explore diverse edit directions. A judge selects a candidate using execution traces and execution-derived proxy signals. The selected harness determines the current batch's outputs and initializes adaptation on the next. Crucially, the solver, proposers, and judge operate through different harnesses around the same frozen LLM, without weight updates, access to final-answer labels, or a separately trained adaptation model. Across six execution-grounded benchmarks spanning text-to-SQL, competitive programming, software engineering, data-science coding, terminal-based tasks, and agentic tool use, TTHE improves fixed ReAct-style baseline harnesses through persistent, inspectable program adaptation. These results recast test-time adaptation for LLM agents as evolution over executable control programs and identify execution-derived proxy reliability as a central challenge for robust unsupervised agent improvement.
Fengxu Yang, Minghui Liwang, Yang Lu, Bo Han, Xianbin Wang, Seyyedali Hosseinalipour
High-quality annotations are essential for training and evaluating artificial intelligence (AI) systems. Quality control (QC) often acquires additional human or AI judgments before accepting a label. Nevertheless, shared biases can cause correlated judgments, leading conventional methods to mistake repeated agreement for independent evidence of correctness. To address this challenge, we propose JABER (Joint Alias-Preserving Belief for Evidence Routing), a bias-aware sequential QC framework that maintains a joint Bayesian posterior over label correctness and shared biases to guide evidence acquisition and termination decisions. Building on this posterior, we develop a finite-horizon algorithm that selects evidence-acquisition and termination actions to minimize expected QC loss under resource constraints. Our theoretical analysis shows that conditional-independence updating can become arbitrarily confident in an incorrect label under persistent shared bias. We further identify when repeated inspections cannot distinguish a correct label from an incorrect label supported by shared bias, when obtaining a new label can resolve this ambiguity, and why modeling the dependence between label correctness and shared bias can be necessary for QC decisions. Our experiments show that JABER reduces expected QC loss by an average of 4.65% compared with a method maintaining separate label correctness and bias probabilities when shared bias makes repeated inspections unreliable. Further, JABER's finite-horizon planning, which evaluates outcomes of successive QC actions, reduces expected QC loss by an average of 10.05% compared with a one-step lookahead method. In real-API experiments on FEVER, JABER achieves zero labeling errors while reducing QC costs by 2.67% and 16.99% compared with marginal-based belief modeling and one-step lookahead methods, respectively.
Abudukelimu Wuerkaixi, Qizhou Wang, Sen Cui, Wutong Xu, Bo Han, Gang Niu, Masashi Sugiyama, Changshui Zhang
With the growing deployment of large language models (LLMs) across diverse domains, concerns regarding their safety have grown substantially. LLM unlearning has emerged as a pivotal approach to removing harmful or unlawful content while maintaining utility. Despite increasing interest, the challenges of continual unlearning, which is common in real-world scenarios, remain underexplored. Successive unlearning tasks often lead to intensified utility degradation. To effectively unlearn targeted knowledge while preserving LLM utility, it is essential to minimize changes in model parameters by selectively updating those linked to the target knowledge, thereby ensuring other knowledge remains unaffected. Building on the task vector framework, we propose a new method named ALKN (Adaptive Localization of Knowledge Negation), which uses dynamic masking to sparsify training gradients and adaptively adjusts unlearning intensity based on inter-task relationships. Comprehensive experiments across three well-established LLM unlearning datasets demonstrate that our approach consistently outperforms baseline methods in both unlearning effectiveness and utility retention under continual unlearning settings.
Chenxi Liu, Yongqiang Chen, Tongliang Liu, Mingming Gong, James Cheng, Bo Han, Kun Zhang
Revealing hidden causal variables alongside the underlying causal mechanisms is essential to the development of science. Despite the progress in the past decades, existing practice in causal discovery (CD) heavily relies on high-quality measured variables, which are usually given by human experts. In fact, the lack of well-defined high-level variables behind unstructured data has been a longstanding roadblock to a broader real-world application of CD. This procedure can naturally benefit from an automated process that can suggest potential hidden variables in the system. Interestingly, Large language models (LLMs) are trained on massive observations of the world and have demonstrated great capability in processing unstructured data. To leverage the power of LLMs, we develop a new framework termed Causal representatiOn AssistanT (COAT) that incorporates the rich world knowledge of LLMs to propose useful measured variables for CD with respect to high-value target variables on their paired unstructured data. Instead of directly inferring causality with LLMs, COAT constructs feedback from intermediate CD results to LLMs to refine the proposed variables. Given the target variable and the paired unstructured data, we first develop COAT-MB that leverages the predictivity of the proposed variables to iteratively uncover the Markov Blanket of the target variable. Built upon COAT-MB, COAT-PAG further extends to uncover a more complete causal graph, i.e., Partial Ancestral Graph, by iterating over the target variables and actively seeking new high-level variables. Moreover, the reliable CD capabilities of COAT also extend the debiased causal inference to unstructured data by discovering an adjustment set. We establish theoretical guarantees for the CD results and verify their efficiency and reliability across realistic benchmarks and real-world case studies.
Muyang Li, Runze Wu IROOTECH TECHNOLOGY, Xiangyu Zhao City University of Hong Kong, Bo Han, Daoyi Dong University of Technology Sydney, Tongliang Liu University of Sydney and RIKEN
The post-training stage of Large Language Models (LLMs) typically involves Supervised Fine-Tuning (SFT) followed by preference alignment to ensure LLM to generate safe, helpful, and instruction-aligned content. The SFT model critically serves as both the initialization and reference model for subsequent preference alignment. However, an essential yet often neglected question is the optimal selection of the SFT checkpoint for this role. We show that checkpoint selection substantially affects final performance, and that the common practice of choosing the minimum validation-loss checkpoint often fails, due to a fundamental conflict between SFT’s focus on imitation and alignment’s goal of response discriminability. To this end, we propose RewardRank, a simple, effective, training-free metrics for estimating initial implicit alignment between reference model and preference objective. Empirical evidence suggests that, using our selected model as reference can gain up to 67.6% relative increase on length-controlled win rate on the popular Zephyr recipe comparing to baselines.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表