Search arXiv⌕ Search

arXiv subjects

Meng Jiang

Publications and source records attributed to Meng Jiang.

At least 19 recordsLinked to original sources
arXiv2026

Training Negatives Still Matter After Calibrating Answerability Readouts

Answerability readouts predict whether a passage supplies an answer, but whether calibration can compensate for removing negative passages during training remains unclear. We present a controlled study of how training-negative construction affects calibrated prediction, ranking, and downstream answer retention. We compare native training, which retains negative passages, with empty training, which removes them, holding questions, labels, positive examples, and model weights fixed. Both are calibrated on the same labeled native examples. On SQuAD 2.0 and RepLiQA, native-trained logistic readouts at selected depths achieve lower Brier loss for Qwen2.5-7B-Instruct and Gemma-2-9B-it, with or without output scores. Native fusion also outperforms calibrated output alone. Post-hoc AUROC analysis shows better ranking under native training, beyond what threshold adjustment alone can explain. On the official SQuAD 2.0 development set, native readout policies improve F1 over empty policies but retain more answers to unanswerable questions. The advantage varies by estimator and depth, while RepLiQA's text shortcuts limit claims about contextual understanding. Unfinished generations prevent complete answer-level evaluation, although worst-case bounds preserve the native-empty F1 gains. Our findings show that training negatives affect ranking and downstream decisions even after common calibration, without establishing a universal advantage for native training.

cs.CL↗
arXiv2026

Strength Controls Change How Interventions Compare Across Sparse Autoencoder Codes

Changing sparse autoencoder (SAE) codes can alter both intervention direction and magnitude, so comparing intervention effects requires a strength control specifying what is held equal. Norm matching and output calibration provide different controls, but whether interventions yield different effects across codes after strength matching remains unclear. We compare intervention effects under these controls for fixed feature groups in Gemma and Llama and observe: (1) Whole-patch scaling. With one coefficient per input, full-output Kullback-Leibler (KL) calibration leaves smaller mean effect gaps from the encoder-based reference intervention than total-norm matching for nonnegative least-squares (NNLS) refits. On new Llama inputs, total-norm matching loses 11 reference transfers, whereas full-output calibration preserves every transfer decision. (2) Positionwise scaling. On a small matched gender-agreement subset, positionwise norm matching leaves smaller observed mean gaps than full-output calibration. (3) Answer exclusion. Compared with unscaled NNLS, answer-excluded calibration reduces gaps in the measured subject-verb agreement comparisons but increases them for gender refits. Full-output KL calibration also reduces larger gaps for norm-matched constructed candidates, although matching coverage is incomplete in several settings. Since calibration uses the evaluated input's outputs, it serves as a diagnostic of residual disagreement.

cs.LG↗
arXiv2026

Uncovering Cross-Objective Interference in Multi-Objective Alignment

We study a persistent failure mode in multi-objective alignment for large language models (LLMs), in which scalarized training improves only some objectives while the others degrade. We formalize this phenomenon as cross-objective interference and, to our knowledge, conduct the first systematic study of scalarization algorithms for multi-objective LLM alignment. The study shows that interference is pervasive across algorithms yet strongly model-dependent. To understand how interference arises, we derive a local covariance law stating that an objective improves or degrades at first order according to the sign of the covariance between its reward and the scalarized score. We extend this law to the clipped surrogate objectives of modern reinforcement fine-tuning and show that it still holds under mild conditions. Building on this law, we propose COVariance-floor Enforced Reweighting (COVER), a one-sided controller that raises an objective's weight only when the covariance between its reward and the clipped advantage weight falls below a target. Through extensive experiments, we find that COVER can mitigate cross-objective interference while matching linear scalarization when objectives already co-improve. Finally, to explain why interference is model-dependent, we complement the local covariance law with a global convergence analysis. This analysis gives sufficient conditions for the non-convex scalarized objective to satisfy the Polyak--Łojasiewicz condition and relates interference to model geometry.

cs.CL↗
arXiv2026

Learning from Runtime Feedback through Failure-Bank Self-Evolution for Vision-Language-Action Models

Vision-language-action (VLA) models generalize broadly across robotic manipulation tasks, but complex environments require balancing task success with unintended contact. Runtime shields can correct individual actions, but they leave the underlying policy unchanged, so repeated disagreements may create a persistent policy-shield mismatch that blocks task progress. To address this challenge, we introduce FailBank, a four-stage self-evolving framework that converts runtime feedback into persistent policy improvement. During collection, a fixed CBF-based safety module serves as an observe-only teacher, producing counterfactual corrections while the policy remains in control. Outcome-aware admission then converts useful proposals into corrective targets and retains successful uncorrected actions as quiet anchors for guarded LoRA updates. We evaluate FailBank on the VLA-Arena benchmark across two difficulty levels and two VLA backbones. Compared with the base policies, FailBank improves the joint success-cost operating point. Across the two backbones, FailBank improves task success rate by 8.5 and 6.9 percentage points, while reducing policy-induced cumulative cost by 35.6\% and 23.8\%, respectively. Compared with runtime shielding, FailBank raises task success rate by 25.4 and 9.5 percentage points, while maintaining comparable policy-induced cumulative cost. These results show that runtime feedback can serve as persistent policy supervision rather than only as a temporary action constraint.

cs.RO↗
arXiv2026

MoFlow: Multi-Objective Agentic Workflow Generation

We study the generation of agentic workflows that jointly optimize multiple objectives, such as accuracy, cost, latency, robustness, and consistency. Existing methods for workflow generation typically optimize accuracy alone or a weighted sum of objectives, so each trained generator commits to one fixed trade-off and must be retrained from scratch when preferences change. To alleviate this, we propose MoFlow, which generates workflows optimized across varied preferences. Specifically, MoFlow formulates workflow generation as a multi-objective Markov decision process and solves it by leveraging Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups, where every node stores a set of reachable trade-offs rather than one weighted score. A single search thus approximately covers the Pareto front, from which MoFlow can return a workflow for any preference by lookup without retraining. We evaluate MoFlow against six strong baselines on six benchmarks spanning mathematics, code, and question answering. Since the baselines are single-scalar optimizers by design, an apples-to-apples comparison is difficult. We instead adopt an evaluation setup that favors the baselines, in that they are rerun for each testing preference, which MoFlow never sees. Even under this stringent setup, MoFlow achieves the highest average hypervolume.

cs.AI↗
arXiv2026

Agent as Policy for Robotic Manipulation

We demonstrate that a general-purpose agent can directly drive a physical robot throughout task execution without any task-specific or environment-specific training. We introduce Agent as Policy (AGP), which places task planning and execution under the agent's control. Given a task and a robot interface, the agent interprets visual evidence, writes executable programs, issues motion commands, and revises its actions in response to physical outcomes. This brings the agent's reasoning and programming capabilities into continuous interaction with the physical world. We study AGP across multiple real-world manipulation tasks spanning precision manipulation, dynamic motions, and deformable objects. These include assembly from human videos, block construction from goal images, dice flipping, targeted throwing, and bimanual towel folding. AGP achieves success rates of at least 80% in seven of eight task configurations and significantly outperforms previous agentic robot systems. We further study efficiency through task experience accumulation and find that reusing saved procedures and programs shortens execution time across repeated trials. These findings support a path for general-purpose agents to act as robot policies, extending their autonomy to physical manipulation through runtime reasoning, programming, and interaction.

cs.CL↗
arXiv2026

Graph Forward Distribution Matching for Molecular Inverse Design

Achieving precise control over multiple properties without sacrificing chemical validity remains a central challenge in molecular inverse design. Existing reinforcement learning (RL) methods fine-tune graph diffusion models by treating **reverse** sampling as a sequential policy, using a single terminal reward to optimize hundreds of coupled decisions. They often suffer from instability, validity collapse, and limited property gains. We introduce GraphFDM (Graph Forward Distribution Matching), a new online RL paradigm for graph diffusion that performs optimization through the **forward** process. GraphFDM uses valid generations to define a reward-tilted target distribution jointly optimized over graph size and molecular structure for each property condition, incorporating reinforcement signals into supervised learning without storing reverse trajectories. We derive the unique optimal target, prove a condition-wise improvement guarantee, and show that the fixed graph-size prior of standard graph diffusion leaves an irreducible matching gap. In multi-conditional polymer and small-molecule generation, GraphFDM achieves the lowest MAE on every target property, with reductions of up to 53.0\% relative to the strongest baselines and chemical validity above 0.99. It further generalizes to out-of-distribution property combinations.

cs.LG↗
arXiv2026

POPI: Personalizing LLMs via Optimized Natural Language Preference Inference

Large language models (LLMs) are typically aligned with population-level preferences, despite substantial variation across individual users. We introduce POPI, a user-level personalization framework that separates the problem into two components connected by a natural-language interface: a shared inference model that distills heterogeneous user signals into a concise preference summary, and a shared generator that conditions on this summary to produce personalized responses. Both components are trained under a unified preference-optimization objective, with reinforcement learning handling the non-differentiable inference step. This objective decomposes into generator approximation error and summary informativeness, revealing how a single loss simultaneously drives accurate generation and informative summarization. Because the interface is natural language, learned summaries can be inferred once per user and reused across different generators -- including frozen, black-box commercial APIs. Across four personalization benchmarks, POPI generally improves personalization quality while reducing context overhead by up to an order of magnitude.

cs.CL↗
arXiv2026

ClinicalReTrial: Clinical Trial Redesign with Self-Evolving Agents

Clinical trials constitute a critical yet exceptionally challenging and costly stage of drug development (\$2.6B per drug), where protocols are encoded as complex natural language documents, motivating the use of AI systems beyond manual analysis. Existing AI methods accurately predict trial failure, but do not provide actionable remedies. To fill this gap, this paper proposes ClinicalReTrial, a multi-agent system that formulates clinical trial optimization as an iterative redesign problem on textual protocols. Our method integrates failure diagnosis, safety-aware modifications, and candidate evaluation in a closed-loop, reward-driven optimization framework. Serving the outcome prediction model as a simulation environment, ClinicalReTrial enables low-cost evaluation and dense reward signals for continuous self-improvement. We further propose a hierarchical memory that captures iteration-level feedback within trials and distills transferable redesign patterns across trials. Empirically, ClinicalReTrial turns 56.7% of failed protocols into predicted successes under the simulation environment, with a mean success probability gain of 7.4% at negligible cost (\$0.156 per trial). Extensive retrospective case studies further show that ClinicalReTrial recovers clinically meaningful modification patterns that align with real-world expert-driven protocol changes.

cs.AI↗
arXiv2026

Steering Multimodal Large Language Models Decoding for Context-Aware Safety

Multimodal Large Language Models (MLLMs) are increasingly deployed in real-world applications, yet their ability to make context-aware safety decisions remains limited. Existing methods often fail to balance oversensitivity (unjustified refusals of benign queries) and undersensitivity (missed detection of visually grounded risks), leaving a persistent gap in safety alignment. To address this issue, we introduce Safety-aware Contrastive Decoding (SafeCoDe), a lightweight and model-agnostic decoding framework that dynamically adjusts token generation based on multimodal context. SafeCoDe operates in two stages: (1) a contrastive decoding mechanism that highlights tokens sensitive to visual context by contrasting real and Gaussian-noised images, and (2) a global-aware token modulation strategy that integrates scene-level reasoning with token-level adjustment to adapt refusals according to the predicted safety verdict. Extensive experiments across diverse MLLM architectures and safety benchmarks, covering undersensitivity, oversensitivity, and general safety evaluations, show that SafeCoDe consistently improves context-sensitive refusal behaviors while preserving model helpfulness.

cs.CL↗
arXiv2026

Knowledge-Verified Emergent Deception in LLM Agents Under Conflicting Incentives

Large language models are increasingly deployed as autonomous agents serving users on behalf of companies, placing them in settings where user and deployer interests can conflict. When an agent knows that a user is owed something its deployer would prefer to deny, does it remain honest? Answering this is difficult because false statements can reflect either ignorance or hallucination rather than deception. To address this challenge, we introduce KnownLieBench , a knowledge-verified benchmark that first confirms through a neutral probe that an agent knows a user's entitlement, and then evaluates whether it makes false claims once an incentive to deny that entitlement is introduced. Specifically, KnownLieBench covers eight customer-service domains and 112 grounded cases, conducts multi-round dialogues with a trust-tracking customer agent, and separates deception emerging from incentive alone from deception produced under explicit instruction. Across eighteen proprietary and open-weight models, emergent deception varies substantially across model families and domains. We further use the benchmark for post-training, finding that honesty-directed fine-tuning reduces deception under incentive, while deception-graded fine-tuning increases lie success on honest-control dialogues without increasing lie frequency under incentive. By verifying entitlement knowledge before scoring deceptive behavior, KnownLieBench reduces the confound between lying and not knowing and enables more rigorous auditing and steering of agent honesty.

cs.CL↗
arXiv2026

ChemDIRT: A Diversified Instruction, Representation, and Task Benchmark for Robust Chemistry-LLM Evaluation

The rapid advancement of large language models (LLMs) has led to increasing interest in their application to scientific domains such as chemistry. However, existing chemistry benchmarks often provide only a narrow view of model capability, focusing on limited task sets while overlooking robustness to variations in problem formulation and chemical representation. As a result, reported performance may overestimate a model's true ability to reason consistently across realistic settings. To address this challenge, we introduce ChemDIRT (Diversified Instruction, Representation, and Task Benchmark), a comprehensive evaluation framework designed to assess the robustness of chemical reasoning in LLMs. ChemDIRT systematically measures model performance across variations in instructions and molecular representations while spanning eight categories of chemistry tasks. By evaluating both accuracy and consistency under these controlled perturbations, ChemDIRT provides a more reliable assessment of model reasoning capabilities than conventional single-format benchmarks. We benchmark a diverse set of open- and closed-source LLMs, revealing substantial prompt sensitivity, representation dependence, and uneven performance across task families.

cs.LG↗
arXiv2026

Forgotten in Weights, Recovered by Tools: Agentic Tool Unlearning for LLM Agents

Large language models (LLMs) are increasingly deployed as tool-augmented agents, where responses can depend on tool calls and external observations rather than model parameters alone. This creates an evaluation mismatch for LLM unlearning: previous unlearning methods may suppress direct parametric recall, but an agent can still recover the same forget target through tools such as web search, retrieval, or database lookup. We identify this failure mode as tool-mediated recovery and study agentic tool unlearning, which aims to reduce both parametric recall and tool-mediated recovery while preserving normal tool use for retained knowledge. To address this challenge, we propose Agentic Tool Unlearning (ATU), a two-stage framework. The first stage applies parametric knowledge unlearning to suppress direct recall, while the second stage performs trajectory-level reinforcement learning in simulated tool-augmented environments to penalize target-seeking tool behavior and final-answer leakage. Experiments on RWKU and MUSE across different LLM architectures show that ATU achieves a better balance between target forgetting and retained utility, making unlearning more robust under tool-augmented agent deployment.

cs.CL↗
arXiv2026

Open, Reliable, and Collective: A Community-Driven Framework for Tool-Using AI Agents

Tool-integrated LLMs retrieve information, perform computations, and take real-world actions, but their reliability depends on both tool-use accuracy and intrinsic tool accuracy, including tool correctness, stability, and safety. While prior work primarily emphasizes tool use, intrinsic tool accuracy remains underexamined. We introduce OpenTools, a community-driven and maintainable toolbox for discovering, using, evaluating, and contributing open-source tools. OpenTools standardizes tool interfaces, converts documented Python functions into reviewable bundles, supports maintainer-triggered evaluation, and combines non-executing risk inspection with optional advisory LLM review. A public web demo allows users to run tools and agents, inspect evidence, contribute tests, and submit tools for maintainer review, while MCP enables controlled access from external applications. Experiments show that community-contributed, task-specific tools yield relative gains of 6% to 22% over an existing toolbox across multiple agent architectures, highlighting the importance of intrinsic tool accuracy.

cs.AI↗
arXiv2026

CreativityPrism: A Cross-Domain Evaluation Framework for Large Language Model Creativity

Creativity is often seen as a hallmark of human intelligence. While large language models(LLMs) are increasingly perceived as generating creative text, there is still no cross-domain and scalable framework to evaluate their creativity across diverse scenarios. Existing methods of LLM creativity evaluation either heavily rely on humans, limiting speed and scalability, or are fragmented across different domains and different definitions of creativity. To address this gap, we propose CreativityPrism, an evaluation and analysis framework that consolidates eight tasks from three domains: divergent thinking, creative writing, and logical reasoning, into a taxonomy of creativity that emphasizes three dimensions: quality, novelty, and diversity of LLM generations. The framework is designed to be scalable with reliable automatic evaluation judges that have been validated against human annotations. We evaluate 17 state-of-the-art (SoTA) LLMs on CreativityPrism and find that while frontier-scale LLMs dominate creative writing and logical reasoning tasks by a .10 (or 15%) lead over locally-deployable open models, they offer no significant advantage in divergent thinking, a domain much less explored in existing post-training regimes. Our analysis also shows that high performance in one creative dimension or domain rarely generalizes to others; specifically, novelty metrics often show weak or negative correlations with other metrics. This fragmentation confirms that a cross-domain, multi-dimensional framework like CreativityPrism is essential for any meaningful assessment of LLM creativity.

cs.CL↗
arXiv2026

CNN Models for Microphone Array Covariance Matrix Upsampling and Acoustic Imaging

Acoustic imaging visualization is a core methodology in acoustics, enabling spatial analysis of sound sources and acoustic scenes. However, limited sensor availability in practical systems motivate approaches that enhance spatial resolution without increasing the hardware complexity. In this paper, we focus on upsampling virtually a tetrahedral 4-microphone array to a spherical 32-microphone array by estimating the covariance matrices of the channels employing deep learning techniques. Five neural network architectures are investigated for covariance upsampling for acoustic imaging using the real-world STARSS23 dataset. These models are developed to estimate a 32-microphone, time-frequency covariance matrix from a 4-microphone input covariance representation. The proposed architectures are based on 2D convolutional layers to capture the underlying spatial-spectral structure of covariance matrices, and are further enhanced with frequency dynamic convolution to model their frequency-dependent properties. The proposed architectures are evaluated in terms of root mean square error (RMSE) and using delay-and-sum beamforming acoustic imaging. Quantitative results show that all models outperform a random-guess baseline, which yields an RMSE of 0.548, with the best-performing architecture achieving an RMSE of 0.432. We analyze qualitatively the performance of the proposed models through beamforming heatmap visualizations derived from the 4-channel input covariance, the 32-channel ground truth, and the predicted 32-channel covariance matrices. These results demonstrate that covariance upsampling significantly enhances the effective performance of the 4-channel microphone array, producing sound maps that closely resemble those obtained with the 32-channel array.

eess.AS↗
arXiv2026

ReplicatorBench: Benchmarking LLM Agents for Replicability in Social and Behavioral Sciences

The literature has witnessed an emerging interest in AI agents for automated assessment of scientific papers. Existing benchmarks focus primarily on the computational aspect of this task, testing agents' ability to reproduce or replicate research outcomes when having access to the code and data. This setting, while foundational, (1) fails to capture the inconsistent availability of new data for replication as opposed to reproduction, and (2) lacks ground-truth diversity by focusing only on reproducible papers, thereby failing to evaluate an agent's ability to identify non-replicable research. Furthermore, most benchmarks only evaluate outcomes rather than the replication process. In response, we introduce ReplicatorBench, an end-to-end benchmark, including human-verified replicable and non-replicable research claims in social and behavioral sciences for evaluating AI agents in research replication across three stages: (1) extraction and retrieval of replication data; (2) design and execution of computational experiments; and (3) interpretation of results, allowing a test of AI agents' capability to mimic the activities of human replicators in real world. To set a baseline of AI agents' capability, we develop ReplicatorAgent, an agentic framework equipped with necessary tools, like web search and iterative interaction with sandboxed environments, to accomplish tasks in ReplicatorBench. We evaluate ReplicatorAgent across four underlying large language models (LLMs), as well as different design choices of programming language and levels of code access. Our findings reveal that while current LLM agents are capable of effectively designing and executing computational experiments, they struggle with retrieving resources, such as new data, necessary to replicate a claim. All code and data are publicly available at https://github.com/CenterForOpenScience/llm-benchmarking.

cs.AI↗
arXiv2026

Controllable Molecular Generative Foundation Models

Despite the success of foundation models in language and vision, molecular graph generation still lacks a unified framework for heterogeneous design tasks with reliable controllability. While reinforcement learning (RL) offers a natural post-training mechanism for task-specific optimization, applying it to graph generative models is hindered by the vast atom-wise action spaces and chemically invalid intermediate states. We propose \textbf{Co}ntrollable \textbf{Mole}cular Generative Foundation Models (CoMole), built with a unified motif-aware graph diffusion pipeline. By learning a motif-aware graph space, CoMole transfers pretrained structural priors into controllable generation, where RL optimizes conditional reverse policies over chemically meaningful decisions. We theoretically characterize the bottleneck of atom-level RL and justify motif-aware policy optimization. Across three heterogeneous benchmarks spanning materials and drug discovery, CoMole ranks first in controllability on all nine targets, reduces MAE by up to 48.2% relative to the strongest baselines, and maintains validity above 0.94 without rule-based correction or post-hoc filtering. We further show that CoMole transfers controllability to unseen properties by optimizing only task embeddings with the generator frozen, achieving performance competitive with strong task-specific baselines.

cs.LG↗