arXiv subjects

Gangyi Zhang

Publications and source records attributed to Gangyi Zhang.

10 recordsLinked to original sources

arXiv2026

Learning Calibrated Confidence for Multi-Agent Collaboration

Complex information seeking on the Web often requires collecting evidence from multiple documents and reasoning across them, as in multi-hop question answering. Multi-agent systems address such tasks by dividing work among specialized agents, where upstream agents produce intermediate answers and supporting evidence that downstream agents build upon. However, upstream answers vary in correctness, and most systems provide no quantitative signal about it, leaving the downstream agent to judge whether to trust or verify what it receives based solely on the textual evidence. In this paper, we propose MASCAL. Each agent learns to produce a correctness probability for its answer by jointly optimizing task performance and confidence calibration in reinforcement learning; the upstream agent passes this probability at handoff, giving the downstream agent a quantitative basis for its decisions, and the final answer also carries a calibrated probability, quantifying the reliability of the system output. To guide the downstream agent in using this signal during training, we introduce confidence-aware weighting, which increases the learning weight at positions responsive to the upstream probability, so that the agent learns more from decisions where it actively uses the signal. Across three collaboration architectures spanning Web retrieval and reasoning tasks, MASCAL improves task accuracy and calibration quality for both agents, and generalizes to deeper collaboration chains and different model scales. Intervention analysis confirms that upstream confidence affects downstream decision behavior, and the learned final confidence supports selective answering and targeted rechecking.

cs.LG↗

arXiv2026

Locating and Closing Information Gaps in Multi-Agent Systems

Multi-step Web tasks often require several LLM agents to gather and combine evidence. The communication graph of a fixed workflow restricts which agents can read each other's outputs: evidence retrieved by one agent may never reach the agent that needs it, leaving an information gap and an incomplete or incorrect answer. Adding a connection can deliver the missing evidence, but can also lead the receiver to adopt an incorrect upstream conclusion that displaces a correct answer; the final output alone does not separate the two. We propose Selective Probing and Link Insertion for Closing information gaps in multi-agent Execution (SPLICE), which locates information gaps per query and decides what to adopt after adding an edge. After the original workflow runs, SPLICE uses the query and node outputs to locate a possible gap, adds one directed edge from a source to a receiver to deliver the relevant evidence, and re-executes the receiver and affected downstream nodes, producing a candidate answer. The adoption decision reads both the final answers of the two runs and the process changes the edit caused inside the workflow: how far the receiver moved, where its new content came from, and whether the change reached the final answer. A selector trained on these process features chooses between adopting the candidate answer and retaining the original. On five benchmarks, SPLICE raises the task score on Web tool use by 8.0% and the mean score on four reasoning and planning benchmarks by 5.7% relative to the static skeleton, at 1.25 to 1.44$\times$ the skeleton's cost, and transfers zero-shot to four unseen backbones.

cs.AI↗

arXiv2026

EyeVQA: Benchmarking Ophthalmic Vision-Language Models from Recognition to Spatial Grounding

Vision-language models (VLMs) have shown increasing potential for medical image understanding, yet their capabilities in ophthalmic imaging remain insufficiently characterized. Existing ophthalmic datasets are typically designed for individual diseases or specialized tasks, making it difficult to systematically evaluate whether VLMs can move beyond disease recognition toward comparative reasoning and fine-grained spatial grounding. We introduce EyeVQA, a unified visual question answering benchmark for comprehensive evaluation of ophthalmic VLMs. EyeVQA is constructed from 21 available ophthalmic datasets and contains 20,000 question-answer pairs spanning six disease groups and seven question types: Single-Choice, Multi-Select, Variable-Select, True-False, Ranking, Point Location, and Bounding Box. Gold answers are deterministically derived from source-provided diagnoses, severity grades, clinical findings, segmentation masks, bounding boxes, and anatomical landmarks, enabling reproducible evaluation without relying on model-generated annotations. Notably, 44.5% of the questions require reasoning across multiple images, extending evaluation beyond conventional single-image medical VQA. We benchmark fourteen representative general-purpose, scientific, and medically specialized VLMs under a unified zero-shot protocol. The best-performing model only achieves an overall score of 62.8, while substantial gaps remain in spatial grounding and cross-task generalization. These results highlight the limitations of current VLMs in comprehensive ophthalmic visual understanding and establish EyeVQA as a diagnostic benchmark for developing more reliable and spatially grounded ophthalmic multimodal models. The project page is available at https://github.com/PKUTHM/EyeVQA.

cs.CV↗

arXiv2026

When Upstream Messages Override Correct Answers: A Controlled Study of Multi-Agent LLM Collaboration

Multi-agent LLM systems rely on message passing among specialized agents to accomplish complex tasks. However, an upstream agent may provide useful information or an incorrect answer that causes a downstream agent to override a correct answer supported by its own evidence. Prior work has not clearly separated the benefits of communication from the damage caused by incorrect messages. We study this problem with controlled experiments across five benchmarks and five receivers, keeping the downstream task and evidence fixed while comparing answers under three conditions: no message, the upstream agent's original message, or a message with the opposite conclusion. Our experiments reveal three key findings. First, messages often help when the downstream agent would otherwise answer incorrectly. Second, messages can also hurt: when the downstream agent would answer correctly without a message, an incorrect upstream message changes the answer in up to 32% of cases. Third, in 94% of audited harmful cases, the downstream agent copies the upstream's specific wrong answer--a pattern we term answer substitution. Removing unreliable messages recovers part of the lost accuracy, suggesting that communication should be selective based on upstream reliability and the evidence already available to the downstream agent.

cs.AI↗

arXiv2026

Beyond Verbalized Confidence: Calibrating Reasoners with Differentiable Readouts

Reinforcement learning with verifiable rewards (RLVR) trains reasoning models to produce correct answers, but does not ensure that their stated confidence is calibrated. The resulting models are systematically overconfident. Recent methods train calibration inside the RLVR loop by having the model state a numerical confidence alongside its answer, but they all obtain the confidence by sampling it as text. This choice imposes two costs: a sampled confidence introduces variance and in practice collapses to a handful of distinct values, and sampling makes the confidence non-differentiable, forcing the calibration loss through a scalar reward. We propose CREDO (Confidence REaDOut) to replace sampling with a deterministic readout. While RLVR optimizes correctness, CREDO reads the confidence from a dedicated token pair in the model's output distribution and trains it by differentiable regression. CREDO further turns the trained confidence into a signal for accuracy, weighting rollouts by how far confidence and outcome disagree, so that accuracy and calibration improve together. Across mathematical and code reasoning, CREDO attains the best accuracy and calibration, and the gains extend to abstention and selective prediction.

cs.LG↗

arXiv2026

Elastic Horizon: Discovering the Effective Interaction Frontier in Agentic Reinforcement Learning

Scaling the interaction horizon-the maximum number of environment interactions per episode-improves LLM agents on long-horizon tasks, and curriculum-based methods that progressively expand the horizon outperform fixed-horizon alternatives. However, existing schedules are open-loop: they monotonically increase the horizon until a manually specified maximum, with no mechanism to detect when further expansion stops helping. We propose the effective interaction frontier hypothesis: a dynamic boundary beyond which additional interactions yield diminishing returns while cost grows linearly. We then introduce Elastic Horizon, a closed-loop controller that tracks this boundary via the 90th percentile of successful trajectory lengths. On AppWorld and BFCL, fixed-horizon sweeps reveal clear saturation plateaus; Elastic Horizon stabilizes the horizon inside the saturation band from both under- and over-capacity initializations, attains the best success rates across 7B and 14B backbones, and saves up to 25% of per-step trajectory tokens. Our work shifts the paradigm from how to scale interaction horizons to when to stop scaling.

cs.AI↗

arXiv2026

Sequence-aware Large Language Models for Explainable Recommendation

Large Language Models (LLMs) have shown strong potential in generating natural language explanations for recommender systems. However, existing methods often overlook the sequential dynamics of user behavior and rely on evaluation metrics misaligned with practical utility. We propose SELLER (SEquence-aware LLM-based framework for Explainable Recommendation), which integrates explanation generation with utility-aware evaluation. SELLER combines a dual-path encoder-capturing both user behavior and item semantics with a Mixture-of-Experts adapter to align these signals with LLMs. A unified evaluation framework assesses explanations via both textual quality and their effect on recommendation outcomes. Experiments on public benchmarks show that SELLER consistently outperforms prior methods in explanation quality and real-world utility.

cs.IR↗

arXiv2025

Vague Preference Policy Learning for Conversational Recommendation

Conversational recommendation systems (CRS) commonly assume users have clear preferences, leading to potential over-filtering of relevant alternatives. However, users often exhibit vague, non-binary preferences. We introduce the Vague Preference Multi-round Conversational Recommendation (VPMCR) scenario, employing a soft estimation mechanism to accommodate users' vague and dynamic preferences while mitigating over-filtering. In VPMCR, we propose Vague Preference Policy Learning (VPPL), consisting of Ambiguity-aware Soft Estimation (ASE) and Dynamism-aware Policy Learning (DPL). ASE captures preference vagueness by estimating scores for clicked and non-clicked options, using a choice-based approach and time-aware preference decay. DPL leverages ASE's preference distribution to guide the conversation and adapt to preference changes for recommendations or attribute queries. Extensive experiments demonstrate VPPL's effectiveness within VPMCR, outperforming existing methods and setting a new benchmark. Our work advances CRS by accommodating users' inherent ambiguity and relative decision-making processes, improving real-world applicability.

cs.IR↗

arXiv2024

Reformulating Conversational Recommender Systems as Tri-Phase Offline Policy Learning

Existing Conversational Recommender Systems (CRS) predominantly utilize user simulators for training and evaluating recommendation policies. These simulators often oversimplify the complexity of user interactions by focusing solely on static item attributes, neglecting the rich, evolving preferences that characterize real-world user behavior. This limitation frequently leads to models that perform well in simulated environments but falter in actual deployment. Addressing these challenges, this paper introduces the Tri-Phase Offline Policy Learning-based Conversational Recommender System (TCRS), which significantly reduces dependency on real-time interactions and mitigates overfitting issues prevalent in traditional approaches. TCRS integrates a model-based offline learning strategy with a controllable user simulation that dynamically aligns with both personalized and evolving user preferences. Through comprehensive experiments, TCRS demonstrates enhanced robustness, adaptability, and accuracy in recommendations, outperforming traditional CRS models in diverse user scenarios. This approach not only provides a more realistic evaluation environment but also facilitates a deeper understanding of user behavior dynamics, thereby refining the recommendation process.

cs.IR↗

arXiv2020

Interactive Path Reasoning on Graph for Conversational Recommendation

Traditional recommendation systems estimate user preference on items from past interaction history, thus suffering from the limitations of obtaining fine-grained and dynamic user preference. Conversational recommendation system (CRS) brings revolutions to those limitations by enabling the system to directly ask users about their preferred attributes on items. However, existing CRS methods do not make full use of such advantage -- they only use the attribute feedback in rather implicit ways such as updating the latent user representation. In this paper, we propose Conversational Path Reasoning (CPR), a generic framework that models conversational recommendation as an interactive path reasoning problem on a graph. It walks through the attribute vertices by following user feedback, utilizing the user preferred attributes in an explicit way. By leveraging on the graph structure, CPR is able to prune off many irrelevant candidate attributes, leading to better chance of hitting user preferred attributes. To demonstrate how CPR works, we propose a simple yet effective instantiation named SCPR (Simple CPR). We perform empirical studies on the multi-round conversational recommendation scenario, the most realistic CRS setting so far that considers multiple rounds of asking attributes and recommending items. Through extensive experiments on two datasets Yelp and LastFM, we validate the effectiveness of our SCPR, which significantly outperforms the state-of-the-art CRS methods EAR (arXiv:2002.09102) and CRM (arXiv:1806.03277). In particular, we find that the more attributes there are, the more advantages our method can achieve.

cs.IR↗