Search arXiv⌕ Search

arXiv subjects

Bo Han

Publications and source records attributed to Bo Han.

At least 19 recordsLinked to original sources
arXiv2026

Discovering and Reasoning of Causality in the Hidden World with Large Language Models

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.

cs.LG↗
arXiv2026

Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation-Level Enforcement

Reinforcement learning (RL) has substantially improved the reasoning capabilities of vision-language models (VLMs), yet it can induce chain-of-thought (CoT) obfuscation, an operational, non-intentional regime where models achieve high accuracy while producing reasoning traces that are less grounded and harder to monitor. While prior work documents this decay at the behavioral level, the underlying representational correlates of such drift remain poorly understood. In this work, we find that obfuscation is accompanied by a representational shift, where RL-induced optimization makes task-agnostic template features less separable from visually grounded content in the model's activation space. Based on this insight, we propose TAME, an activation-level intervention framework that leverages Sparse Autoencoders (SAEs) to regularize VLM internal features during RL training. Specifically, TAME utilizes an anti-obfuscation feedback loop to reduce behavioral obfuscation and applies an asymmetric SAE constraint to specific internal directions, penalizing template activations while preserving visual evidence. By directly intervening on the features associated with obfuscation, our method improves the transparency and visual grounding of reasoning traces. Experiments on the VIRL-39k and SPA-VL benchmarks across two model families show that TAME improves CoT monitorability (by up to +30.9 points on VIRL and +16.7 points on SPA-VL over GRPO), while largely preserving task accuracy and general-capability performance across four standard VLM benchmarks, with task-specific trade-offs. Our code is available at https://github.com/henrymao2004/TAME.

cs.AI↗
arXiv2026

UnlearningSoup: Is Repeated Tuning Necessary for Large Language Model Unlearning?

Large language models trained on vast corpora inherently risk memorizing harmful content that may later re-emerge in their outputs. To mitigate this issue, existing unlearning methods typically rely on training-based parameter updates, such as gradient ascent and its variants, to delete targeted content while preserving other knowledge. However, balancing the competing goals of forgetting and retention makes hyperparameter choices for these methods particularly difficult, often requiring repeated tuning to obtain a strong model that still leaves substantial room for improvement and transfers poorly across models and datasets. To address this challenge, we investigate whether unlearning runs exhibit exploitable structure in weight space, and observe that models from different runs still lie in a shared evaluation-performance basin. This suggests that stronger models may be recovered through an unlearning-tailored soup strategy, reducing the need for repeated tuning for further improvement or new settings. Motivated by this, we propose UnlearningSoup, a unified framework that provides two strategies: EfficientSoup uses binary-search-based interpolation to quickly discover a well-performing model in the early stage, where repeated tuning would otherwise make strong model selection costly. PerformanceSoup uses reweighted souping to efficiently unlock the remaining performance potential in the later stage, where repeated tuning becomes increasingly inefficient. Extensive experiments across diverse datasets and models show that UnlearningSoup delivers 2.4x to 3.3x efficiency gains in hyperparameter selection, while consistently improving performance across settings.

cs.LG↗
arXiv2026

ExpEvolve: Structuring and Reusing Experience for Open-Ended Discovery

Evolutionary coding agents such as AlphaEvolve are one of the notable successes in using Large Language Models (LLMs) to build AI Scientists. These agents tackle open-ended scientific problems by iteratively improving and evolving programs, leveraging the prior knowledge and reasoning capabilities of LLMs. Despite their success, mutations over programs are mostly low-level upon surface heuristics, without a persistent unit of control. In contrast, science compounds through ideas. In this work, we propose ExpEvolve, which injects idea-level extraction and control into evolution: it discovers and leverages the high-level factors from programs that explain scoring differences among different programs. We characterize theoretical conditions on when discovery with idea factors improves sample efficiency. We also build a practical framework to show that the extracted high-level idea factors are helpful for guiding exploitation from experience and exploration when discovery comes to a plateau. Across 11 open-ended mathematical tasks, ExpEvolve significantly improves the final performance and sample efficiency up to 5.0x. The improvements also transfer to 4 other LLMs and to 6 machine learning tasks from MLS-Bench. Analysis shows that ExpEvolve can discover idea-level factors to provide better exploration and exploitation for evolutionary discovery.

cs.LG↗
arXiv2026

Benchmarking and Error Analysis of Deep Search and Deep Research LLM Agents

As Large Language Models (LLMs) evolve into autonomous Deep Search and Deep Research Agents capable of executing long-horizon, multi-step tasks, evaluating their reliability becomes increasingly critical. However, existing evaluation paradigms predominantly focus on final task outcomes, leaving a significant gap in understanding the intermediate execution traces, where agents are highly susceptible to various errors. To address this critical gap, we introduce a comprehensive methodology and analytical framework designed specifically to diagnose errors in intermediate workflows. Building upon this framework, we establish a structured taxonomy that categorizes these vulnerabilities into two primary dimensions: System-Level Errors and Task-Level Errors. Through a systematic examination of step-by-step execution traces and agent-environment interactions, we reveal how these underlying errors affect overall performance. Finally, through empirical experiments and systematic analysis of the identified error types, we provide actionable design guidelines, aiming to enhance the robustness, reasoning capabilities, and reliability of Deep Search and Deep Research agents.

cs.LG↗
arXiv2026

Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory

To enable reliable long-term interaction, LLM agents require a memory system that can faithfully store, efficiently retrieve, and deeply reason over accumulated dialogue history. Most existing methods adopt an extracted fact based paradigm: handcrafted static prompts compress raw dialogues into atomic facts, which are then stored, matched, and injected into downstream reasoning. Nevertheless, such fact-centric designs inevitably discard fine-grained details in original dialogues and fail to support deep reasoning over scattered isolated facts. Moreover, static prompts cannot maintain consistent extraction granularity across diverse dialogue styles. To address these limitations, we propose TriMem, which maintains three coexisting representation granularities, including raw dialogue segments anchored by source identifiers for storage fidelity, extracted atomic facts for efficient memory retrieval, synthesized profiles that aggregate dispersed facts into holistic semantic understanding for deep reasoning. We further adopt TextGrad-based prompt optimization, which iteratively refines extraction and profiling prompts via response quality feedback, achieving lifelong evolution without any parameter updating. Extensive experiments on LoCoMo and PerLTQA across multiple LLM backbones demonstrate that TriMem consistently outperforms strong memory baselines. The code is available at https://TMLR-TriMem.github.io .

cs.CL↗
arXiv2026

Reclaiming Low-Quality Instruction Data with Rule-Guided LLM Reconstruction

Data selection for supervised fine-tuning (SFT) often discards low-scoring instruction examples that may retain useful supervision. We propose a rule-guided LLM reconstruction framework to reclaim this information. The framework identifies low-scoring samples through LLM-based rating and score correction, selects semantically representative sources, and reconstructs instruction-response pairs through strategy-guided generation, criteria-based validation, and failure-guided repair. From 123,786 low-scoring samples, it constructs 54,888 training examples. Across three 7B/8B base LLMs, the reconstructed data achieves higher average downstream performance than the low-scoring source set and all evaluated data-selection baselines. Despite using substantially fewer examples than the full source dataset (approximately 300K), it achieves higher reported average scores than full-data training across all three base models. Stepwise ablations identify rule-guided reconstruction as the largest source of observed improvements over direct fusion, while a manual audit reveals remaining limitations in source integration and answer faithfulness. These findings support reconstruction as a complementary approach to selection-based SFT data curation.

cs.CL↗
arXiv2026

AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks

Agents pair reasoning models with harnesses that organize model calls, execute tools, load skill instructions, and retain history in memory. Understanding agent performance requires linking execution conditions to outcomes, behavior, and resource use. We introduce AlphaDiana, a measurement system with modular interfaces for datasets, agents, environments, and scorers. It preserves harness-native control loops while controlling file reuse and memory retention separately. AlphaDiana supports macro analyses of complete agents and micro interventions on agent components such as tools, skills, and memory under matched conditions. We study three open-source models and three open-source harnesses on five answer-based benchmarks, with two models also tested on two interactive benchmarks. We find that harnesses do not consistently improve the accuracy of reasoning: gains on questions unsolved by the model alone can be outweighed by losses on questions it answers correctly without tools. This inconsistency extends to component interventions: neither skill guidance nor retained history reliably improves accuracy. Behavior and output-token use must be assessed with correctness: low token entropy can accompany wrong answers, and less output per execution can mean more output tokens per correct answer when accuracy falls. AlphaDiana links answer correctness and execution status to traces and resource records for comparing agents and diagnosing failures under shared task and resource policies.

cs.AI↗
arXiv2026

Understanding Negative Learning in Large Language Model Fine-tuning

Negative learning has never been more important than it is today, particularly in large language model (LLM) fine-tuning, where it underpins machine unlearning, preference- and reward-based optimization, and exploratory data generation. Yet its functional role remains a subject of ongoing debate, with existing work variously characterizing it as forgetting, regularization, or target shaping. However, analytical support focuses primarily on the single-update scale, while empirical results at the trajectory scale are typically obtained under negative-only training and may not extend to composite objectives in which it interacts with other components. To develop a deeper understanding, we introduce a gradient-based endpoint probing framework that can examine negative updates separately throughout training. Our analysis reveals that negative learning plays different roles across algorithms and training phases. For example, in off-policy DPO, it acts more like a regularizer early in training and shows evidence of target shaping only toward out-of-distribution endpoints and only in later phases. In on-policy PPO and GRPO, by contrast, negative updates offset endpoint-aligned positive learning without stable target shaping, consistent with a role in sustaining exploration. Guided by these findings, we develop a family of stage-aware methods that modulate the strength of negative updates throughout training, improving fine-tuning performance across tasks.

cs.LG↗
arXiv2026

Decouple to Forget: Semantic Mediation for Sensitive Unlearning in Large Language Models

Machine unlearning for large language models seeks to remove specified knowledge while preserving general capability, yet current methods often treat an entire response as the unit of forgetting. This coarse granularity causes unnecessary utility loss when the forget target is only a small set of sensitive facts embedded within otherwise benign explanations. In this paper, we introduce Decouple to Forget (D2F), a targeted unlearning method inspired by semantic mediation that separates sensitive factual content from non-sensitive meaning and surface expression. For each example to be forgotten, D2F extracts a fact-level "mediator" representation from the answer, identifies which parts are sensitive, and removes only those parts while leaving the rest of the structure intact. We then generate counterfactual answers grounded in this mediator that preserve non-sensitive content but exclude sensitive facts, employing preference-based contrastive optimization to reduce the likelihood of information leakage while maintaining helpfulness and fluency. Experiments across various benchmark datasets show that D2F achieves a better balance between forgetting and retention than strong baseline methods. Our code is available at https://github.com/JingyueCong/D2F.

cs.CL↗
arXiv2026

Encoding Context into LLM Weights: What Makes It Context?

Longer contexts allow large language models (LLMs) to condition their predictions on more information, but inference costs grow prohibitively with context length. An emerging alternative uses test-time updates to encode context as temporary changes to model weights. However, encoding this information into weights that already represent general knowledge obscures its role as context, so the LLM may fail to use it properly to condition subsequent inference. To restore this contextual role, we first associate each context with an identifier during encoding, thereby distinguishing the newly encoded information from pre-existing knowledge. Invoking the identifier at inference then enables the LLM to recite the encoded context. Nevertheless, successful recitation does not ensure that the LLM will use the recalled information as context. We therefore further introduce a prompting framework that explicitly guides the LLM to use this information as context. Across diverse controlled tasks and practical benchmarks, our method outperformed the baseline using only off-the-shelf LLMs, showing its advances in context encoding.

cs.CL↗
arXiv2026

Measurement-induced criticality for non-Abelian anyons: The monitored golden chain

Non-Abelian anyon chains whose interactions favor a single fusion channel host quantum critical ground states described by minimal-model conformal field theories (CFTs) protected by a topological symmetry. We show that their measurement-only counterpart, where non-commuting fusion charge measurements act locally on the anyons of a chain of $L$ sites, provides a fundamental example of a non-unitary CFT controlling a non-Abelian learning problem. Mapping the monitored dynamics to a random tensor network, we numerically find $1/L$ scaling Lyapunov gaps, logarithmic entanglement scaling with the chord length, and an effective central charge in the Shannon entropy of the measurement record. The conserved topological charge sharpens with dynamical exponent $z= 1$ and is learnable by a decoder on an $O(L)$ time scale. As the topological symmetry is a non-invertible symmetry of the $SU(2)_k$ anyon theories, we establish the monitored golden chain $(k=3)$ as a non-Abelian generalization of the weakly (Kramers-Wannier) self-dual quantum critical point of the toric code $(k=2)$, and the first of a family for $k\geq 3$.

cond-mat.str-el↗
arXiv2026

TTHE: Test-Time Harness Evolution

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.

cs.SE↗
arXiv2026

DRNOISE: Benchmarking Deep Research Agents in Misleading Evidence Environments

Deep research agents increasingly operate over the open web, where relevant records coexist with redundant summaries, outdated reports, and misleading documents. Existing evaluations offer limited insight into whether agents preserve sound evidential standards when an ordinary-looking false document is deliberately seeded into a searchable environment and offers a direct shortcut to a conflicting answer. We introduce DRNOISE, a 100-task benchmark for answer recovery under misleading evidence. Each task has a unique gold answer supported by two corroborating indirect record chains; the paired noisy condition adds one plausible document that states a conflicting answer directly. The benchmark spans ten families of evidence operations. Among models with strong clean-task performance in the shared agent framework, this single intervention causes 66--88 percentage-point accuracy drops. Trace analyses reveal a pattern of verification inertia: agents often retrieve truthful records yet defer to the answer-like document without adequately checking its claim against the underlying evidence. Generic verification prompts reduce but do not close this gap. The setting is especially relevant to open-web deployment, where plausible falsehoods arrive through ordinary-looking pages without explicit adversarial instructions. Reliable deep research therefore requires more than retrieval and citation; it requires active reconciliation of direct claims with record-level evidence.

cs.LG↗
arXiv2026

EvolveNet: Collaborative Harness Evolution for Agent Self-Improvement

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.

cs.LG↗
arXiv2026

Agents Don't Just Agree, They Remember: Benchmarking Persistent Sycophancy in Self-Improving Personal Agents

Self-improving personal agents write profiles, memories, and reusable skills that carry over from one chat to the next. Prior works ask whether a model agrees with the user in its next reply. Yet these agents can also write down a claim the user made, and a later chat can read that note back as trusted context. We call this persistent sycophancy. We introduce the Personal Agent Sycophancy Benchmark, PASB, with 1,600 tasks run on two agent frameworks, Hermes-Agent and OpenClaw, across sixteen models that span scales and generations. Each task separates a first chat containing the claim from a follow-up chat in a fresh session, so the first chat can reach the follow-up only through a note the agent chose to write. Downstream failure, the rate at which follow-up answers still rely on the claim, reaches 68.1% when the agent saved the claim against 35.6% when it did not. Later generations fail less often, yet a saved note still more than doubles their failure rate. In 38.8% of runs the saved note states the claim with more authority than the user gave it, and a saved claim still shapes answers in a different domain. Among four mitigations, editing the stored note works better than correcting the agent in conversation, cutting same-domain downstream failure from 36.3% to 30.3% for Hermes-Agent and from 37.9% to 31.6% for OpenClaw. A strong instruction against saving brings failure back to the level of runs that saved nothing. Self-improving agents need controls on what they write to memory as well as on what they say. Our benchmark is available at https://github.com/henrymao2004/agent-sycophancy.

cs.AI↗
arXiv2026

Exciton-polariton condensate in the van der Waals magnet CrSBr

Van der Waals magnets are an emergent material class of paramount interest for fundamental studies in coupling light with matter excitations, which are uniquely linked to their underlying magnetic properties. Among these materials, the semiconducting magnet CrSBr is possibly a first playground where we can study simultaneously the interaction of photons, magnons, and excitons at the quantum level. Here we demonstrate a coherent macroscopic quantum phase, the bosonic condensation of exciton-polaritons, emerging in a CrSBr flake embedded in a fully tunable cryogenic open optical cavity. The Bose condensate is characterized by a highly non-linear threshold-like behavior, macroscopic occupation of the ground state and coherence manifests distinctly via its first and second order quantum correlations. We find that the condensate's non-linearity is highly susceptible to the magnetic order in CrSBr. Specially, it can encounter a sign change from attractive to repulsive interactions when the intrinsic antiferromagnetic order transforms to the forced ferromagnetic order. Our findings open a route towards magnetically controllable quantum fluids of light, and optomagnonic devices where spin magnetism is coupled to on-chip Bose-Einstein condensates.

cond-mat.mtrl-sci↗
arXiv2026

Rethinking Faithfulness in LLMs: A Pairwise Context-Sensitive Perspective

Large language models (LLMs) are expected to answer questions faithfully based on the provided context, abstaining when the context information is insufficient to answer the questions. Existing faithfulness evaluations typically assess each question-context instance in isolation; however, such instance-level evaluation fails to capture a fundamental requirement of faithful behavior: the ability to adapt model responses to changes in available contexts. In particular, a model should provide correct answers when sufficient evidence is present and abstain when it is not. In this work, we propose a Pairwise Faithfulness Benchmark (PFaithBench) that evaluates whether a model can switch between answering and abstaining for the same question under supporting versus non-supporting contexts. Our evaluations across thirty-nine models with seven model families demonstrate that faithfulness fundamentally involves a trade-off between answering and abstaining, and that most current models exhibit a strong bias toward answering, with most faithfulness errors arising from over-answering, i.e., models tend to fabricate a response even when the provided context is insufficient. We further conduct a series of studies on faithfulness training under different data constructions. Our results show that training outcomes are highly sensitive to the specific composition of answering and abstaining data. Constructing answering and abstaining data from mismatched sources can cause models to rely on dataset-specific shortcuts rather than actual context sufficiency. Moreover, increasing answer-supervised data improves answering performance but exacerbates over-answering, while increasing abstaining data reduces hallucination but leads to over-abstention. The code and data are released at https://github.com/tmlr-group/PFaithBench.

cs.CL↗