arXiv subjects

Qi Cao

Publications and source records attributed to Qi Cao.

At least 37 records · Page 2Linked to original sources

arXiv2026

Self-Harmony: Learning to Harmonize Self-Supervision and Self-Play in Test-Time Reinforcement Learning

Test-time reinforcement learning (TTRL) offers a label-free paradigm for adapting models using only synthetic signals at inference, but its success hinges on constructing reliable learning signals. Standard approaches such as majority voting often collapse to spurious yet popular answers. We introduce Self-Harmony, a framework built on a simple intuition: the correct answer should remain stable across both an original question and its paraphrase. Self-Harmony operationalizes this by employing a single model in two complementary roles: a Solver to produce answers and a Reframer to rephrase the input. Based on this, we further propose a pseudo-label method: instead of majority voting, it aggregates answer frequencies across these original and reframed views using the harmonic mean. This is a process that naturally selects for solutions stable under reframing, thereby avoiding the common trap of favoring view-dependent, spurious answers. Crucially, this requires no human supervision or auxiliary models. Across diverse reasoning benchmarks, Self-Harmony achieves state-of-the-art results at the label-free test-time setting, ranking first in 28 of 30 settings across multiple methods. Beyond accuracy, it demonstrates unprecedented robustness, with zero training failures in all experiments, underscoring its stability and reliability.

cs.CL↗

arXiv2026

Fine-tuning Done Right in Model Editing

Fine-tuning, a foundational method for adapting large language models, has long been considered ineffective for model editing. Here, we challenge this belief, arguing that the reported failure arises not from the inherent limitation of fine-tuning itself, but from adapting it to the sequential nature of the editing task, a single-pass depth-first pipeline that optimizes each sample to convergence before moving on. While intuitive, this depth-first pipeline coupled with sample-wise updating over-optimizes each edit and induces interference across edits. Our controlled experiments reveal that simply restoring fine-tuning to the standard breadth-first (i.e., epoch-based) pipeline with mini-batch optimization substantially improves its effectiveness for model editing. Moreover, fine-tuning in editing also suffers from suboptimal tuning parameter locations inherited from prior methods. Through systematic analysis of tuning locations, we derive LocFT-BF, a simple and effective localized editing method built on the restored fine-tuning framework. Extensive experiments across diverse LLMs and datasets demonstrate that LocFT-BF outperforms state-of-the-art methods by large margins. Notably, to our knowledge, it is the first to sustain 100K edits and 72B-parameter models,10 x beyond prior practice, without sacrificing general capabilities. By clarifying a long-standing misconception and introducing a principled localized tuning strategy, we advance fine-tuning from an underestimated baseline to a leading method for model editing, establishing a solid foundation for future research.

cs.CL↗

arXiv2026

ConFoThinking: Consolidated Focused Attention Driven Thinking for Visual Question Answering

Thinking with Images improves fine-grained VQA for MLLMs by emphasizing visual cues. However, tool-augmented methods depend on the capacity of grounding, which remains unreliable for MLLMs. In parallel, attention-driven methods to crop the Region of Interest (ROIs) are proposed but they are constrained by (1) fragmented attention signals scattered across layers, leading to suboptimal localization and (2) relying on question- or redundant-text-conditioned attention extraction. Our analysis reveals three patterns: MLLMs may attend to the correct region yet generate incorrect coordinates, where-to-look attention is often fragmented across layers, and attention extraction is query-sensitive. Motivated by these, We propose ConFoThinking, a Consolidated-Focused-Attention-Driven Thinking framework that learns to aggregate attention into a designated intermediate layer, from which we mine and zoom in salient regions for downstream visual understanding. Moreover, we extract attention using concise semantic cues of what to look into, which mitigates the semantic noise introduced by question- or redundant-text-based attention extraction. Experiments across five VQA benchmarks demonstrate ConFoThinking significantly improves perception performance. The code, checkpoints, and dataset will be released after being accepted.

cs.CV↗

arXiv2026

Models Under SCOPE: Scalable and Controllable Routing via Pre-hoc Reasoning

Model routing chooses which language model to use for each query. By sending easy queries to cheaper models and hard queries to stronger ones, it can significantly reduce inference cost while maintaining high accuracy. However, most existing routers treat this as a fixed choice among a small set of models, which makes them hard to adapt to new models or changing budget constraints. In this paper, we propose SCOPE (Scalable and Controllable Outcome Performance Estimator), a routing framework that goes beyond model selection by predicting their cost and performance. Trained with reinforcement learning, SCOPE makes reasoning-based predictions by retrieving how models behave on similar problems, rather than relying on fixed model names, enabling it to work with new, unseen models. Moreover, by explicitly predicting how accurate and how expensive a model will be, it turns routing into a dynamic decision problem, allowing users to easily control the trade-off between accuracy and cost. Experiments show that SCOPE is more than just a cost-saving tool. It flexibly adapts to user needs: it can boost accuracy by up to 25.7% when performance is the priority, or cut costs by up to 95.1% when efficiency matters most.

cs.LG↗

arXiv2026

DAJ: Data-Reweighted LLM Judge for Test-Time Scaling in Code Generation

Test-time scaling for code generation commonly relies on Best-of-N selection, in which multiple candidate solutions are sampled from a base model, and the best one is selected by an LLM judge. However, training reliable LLM judges is challenging due to severe distribution shifts, including imbalances between easy and hard problems, mismatches between training tasks and evaluation benchmarks, and trajectory mismatch arising from training data generated by cheaper models whose behavior differs from that of inference-time models. We propose DAJ, a reasoning-based LLM judge trained with verifiable rewards under a bi-level data-reweighted learning framework. The proposed framework learns data-importance weights (either domain-level or instance-level) to optimize generalization performance on a held-out meta set aligned with target benchmarks. To the best of our knowledge, this is the first application of data reweighting to LLM-as-a-Judge training for test-time scaling. Our approach automatically emphasizes hard problems, in-distribution samples, and trajectory-aligned data, without relying on hand-crafted heuristics. Empirically, DAJ achieves state-of-the-art performance on LiveCodeBench and BigCodeBench, outperforming strong test-time scaling baselines as well as leading proprietary models.

cs.LG↗

arXiv2026

FunPRM: Function-as-Step Process Reward Model with Meta Reward Correction for Code Generation

Code generation is a core application of large language models (LLMs), yet LLMs still frequently fail on complex programming tasks. Given its success in mathematical reasoning, test-time scaling approaches such as Process Reward Model (PRM)-based Best-of-N selection offer a promising way to improve performance. However, existing PRMs remain ineffective for code generation due to the lack of meaningful step decomposition in code and the noise of Monte Carlo-estimated partial-solution correctness scores (rewards). To address these challenges, we propose FunPRM. FunPRM prompts LLMs to encourage modular code generation organized into functions, with functions treated as PRM reasoning steps. Furthermore, FunPRM introduces a novel meta-learning-based reward correction mechanism that leverages clean final-solution rewards obtained via a unit-test-based evaluation system to purify noisy partial-solution rewards. Experiments on LiveCodeBench and BigCodeBench demonstrate that FunPRM consistently outperforms existing test-time scaling methods across five base LLMs, notably achieving state-of-the-art performance on LiveCodeBench when combined with O4-mini. Furthermore, FunPRM produces code that is more readable and reusable for developers.

cs.LG↗

arXiv2026

GoalRank: Group-Relative Optimization for a Large Ranking Model

Mainstream ranking approaches typically follow a Generator-Evaluator two-stage paradigm, where a generator produces candidate lists and an evaluator selects the best one. Recent work has attempted to enhance performance by expanding the number of candidate lists, for example, through multi-generator settings. However, ranking involves selecting a recommendation list from a combinatorially large space. Simply enlarging the candidate set remains ineffective, and performance gains quickly saturate. At the same time, recent advances in large recommendation models have shown that end-to-end one-stage models can achieve promising performance with the expectation of scaling laws. Motivated by this, we revisit ranking from a generator-only one-stage perspective. We theoretically prove that, for any (finite Multi-)Generator-Evaluator model, there always exists a generator-only model that achieves strictly smaller approximation error to the optimal ranking policy, while also enjoying scaling laws as its size increases. Building on this result, we derive an evidence upper bound of the one-stage optimization objective, from which we find that one can leverage a reward model trained on real user feedback to construct a reference policy in a group-relative manner. This reference policy serves as a practical surrogate of the optimal policy, enabling effective training of a large generator-only ranker. Based on these insights, we propose GoalRank, a generator-only ranking framework. Extensive offline experiments on public benchmarks and large-scale online A/B tests demonstrate that GoalRank consistently outperforms state-of-the-art methods.

cs.IR↗

arXiv2025

DreamPRM-Code: Function-as-Step Process Reward Model with Label Correction for LLM Coding

Process Reward Models (PRMs) have become essential for improving Large Language Models (LLMs) via test-time scaling, yet their effectiveness in coding remains limited due to the lack of meaningful step decompositions in code and the noise of Monte-Carlo-generated partial labels. We propose DreamPRM-Code, a coding-focused PRM that treats functions as reasoning steps using a Chain-of-Function prompting strategy to induce modular code generation, enabling PRM training and application analogous to mathematical reasoning tasks. To address label noise, DreamPRM-Code introduces a meta-learning-based correction mechanism that leverages clean final-solution unit-test labels and performs bi-level optimization to refine intermediate labels. Applying on test-time scaling, DreamPRM-Code achieved state-of-the-art performance on LiveCodeBench with 80.9 pass@1 rate, surpassing OpenAI o4-mini.

cs.LG↗

arXiv2025

The 2nd Workshop on Human-Centered Recommender Systems

Recommender systems shape how people discover information, form opinions, and connect with society. Yet, as their influence grows, traditional metrics, e.g., accuracy, clicks, and engagement, no longer capture what truly matters to humans. The workshop on Human-Centered Recommender Systems (HCRS) calls for a paradigm shift from optimizing engagement toward designing systems that truly understand, involve, and benefit people. It brings together researchers in recommender systems, human-computer interaction, AI safety, and social computing to explore how human values, e.g., trust, safety, fairness, transparency, and well-being, can be integrated into recommendation processes. Centered around three thematic axes-Human Understanding, Human Involvement, and Human Impact-HCRS features keynotes, panels, and papers covering topics from LLM-based interactive recommenders to societal welfare optimization. By fostering interdisciplinary collaboration, HCRS aims to shape the next decade of responsible and human-aligned recommendation research.

cs.IR↗

arXiv2025

DreamPRM: Domain-Reweighted Process Reward Model for Multimodal Reasoning

Reasoning has substantially improved the performance of large language models (LLMs) on complicated tasks. Central to the current reasoning studies, Process Reward Models (PRMs) offer a fine-grained evaluation of intermediate reasoning steps and guide the reasoning process. However, extending PRMs to multimodal large language models (MLLMs) introduces challenges. Since multimodal reasoning covers a wider range of tasks compared to text-only scenarios, the resulting distribution shift from the training to testing sets is more severe, leading to greater generalization difficulty. Training a reliable multimodal PRM, therefore, demands large and diverse datasets to ensure sufficient coverage. However, current multimodal reasoning datasets suffer from a marked quality imbalance, which degrades PRM performance and highlights the need for an effective data selection strategy. To address the issues, we introduce DreamPRM, a domain-reweighted training framework for multimodal PRMs which employs bi-level optimization. In the lower-level optimization, DreamPRM performs fine-tuning on multiple datasets with domain weights, allowing the PRM to prioritize high-quality reasoning signals and alleviating the impact of dataset quality imbalance. In the upper-level optimization, the PRM is evaluated on a separate meta-learning dataset; this feedback updates the domain weights through an aggregation loss function, thereby improving the generalization capability of trained PRM. Extensive experiments on multiple multimodal reasoning benchmarks covering both mathematical and general reasoning show that test-time scaling with DreamPRM consistently improves the performance of state-of-the-art MLLMs. Further comparisons reveal that DreamPRM's domain-reweighting strategy surpasses other data selection methods and yields higher accuracy gains than existing test-time scaling approaches.

cs.LG↗

arXiv2025

Operator-Consistent Physics-Informed Learning for Wafer Thermal Reconstruction in Lithography

Thermal field reconstruction in post-exposure bake (PEB) is critical for advanced lithography, yet current physics-informed neural networks (PINNs) suffer from inconsistent accuracy due to a misalignment between geometric coordinates, physical fields, and differential operators. To resolve this, we introduce a novel architecture that unifies these elements on a single computation graph by integrating LSTM-gated mechanisms within a Liquid Neural Network (LNN) backbone. This specific combination of gated liquid layers is necessary to dynamically regulate the network's spectral behavior and enforce operator-level consistency, which ensures stable training and high-fidelity predictions. Applied to a 2D PEB scenario with internal heat generation and convective boundaries, our model formulates residuals via differential forms and a composite loss functional. The results demonstrate rapid convergence, uniformly low errors, strong agreement with FEM benchmarks, and stable training without late-stage oscillations, outperforming existing baselines in accuracy and robustness. Our framework thus establishes a reliable foundation for high-fidelity thermal modeling and offers a transferable strategy for operator-consistent neural surrogates in other physical domains.

math-ph↗

arXiv2025

DreamPRM-1.5: Unlocking the Potential of Each Instance for Multimodal Process Reward Model Training

Training multimodal process reward models (PRMs) is hard due to (i) distribution shift between training set and test set and (ii) quality imbalance across training data samples. While domain-level reweighting (e.g., DreamPRM) aligns training with test-time objectives, it leaves a clear gap to an oracle upper bound (pass@N), even under a "sanity check" that uses test set data to probe headroom -- pointing to meta-level under-parameterization. We introduce DreamPRM-1.5, an instance-level reweighting framework that assigns an adaptive weight to every training example via bi-level optimization. To realize instance reweighting across scales, we develop two complementary regimes: Instance Table, which learns explicit per-sample weights and excels on small/medium data, and Instance Net, a lightweight neural network that generalizes better and scales to large corpora. A practical, stable training recipe -- time-scale matching between upper/lower updates, cold-start initialization, and bounded-range weights -- prevents divergence. Integrated with test-time scaling, DreamPRM-1.5 attains 84.6 accuracy on the MMMU validation set, 31.3 accuracy on R-Bench-V and, when paired with a leading backbone (e.g., GPT-5-mini), achieves first-place results on public multimodal reasoning leaderboards. Moreover, extensive experiments, including benchmark evaluations, baseline comparisons, and a sanity check, demonstrate that DreamPRM-1.5 closes the gap toward the oracle, achieves leading performance, and trains stably.

cs.LG↗

arXiv2025

Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs

As large language models (LLMs) often generate plausible but incorrect content, error detection has become increasingly critical to ensure truthfulness. However, existing detection methods often overlook a critical problem we term as self-consistent error, where LLMs repeatedly generate the same incorrect response across multiple stochastic samples. This work formally defines self-consistent errors and evaluates mainstream detection methods on them. Our investigation reveals two key findings: (1) Unlike inconsistent errors, whose frequency diminishes significantly as the LLM scale increases, the frequency of self-consistent errors remains stable or even increases. (2) All four types of detection methods significantly struggle to detect self-consistent errors. These findings reveal critical limitations in current detection methods and underscore the need for improvement. Motivated by the observation that self-consistent errors often differ across LLMs, we propose a simple but effective cross-model probe method that fuses hidden state evidence from an external verifier LLM. Our method significantly enhances performance on self-consistent errors across three LLM families.

cs.CL↗

arXiv2025

A Hypergraph Approach to Distributed Broadcast

This paper explores the distributed broadcast problem within the context of network communications, a critical challenge in decentralized information dissemination. We put forth a novel hypergraph-based approach to address this issue, focusing on minimizing the number of broadcasts to ensure comprehensive data sharing among all network users. The key contributions of this work include the establishment of a general lower bound for the problem using the min-cut capacity of hypergraphs, and a distributed broadcast for quasi-trees (DBQT) algorithm tailored for the unique structure of quasi-trees, which is proven to be optimal. This paper advances both network communication strategies and hypergraph theory, with implications for a wide range of real-world applications, from vehicular and sensor networks to distributed storage systems.

cs.IT↗

arXiv2025

A Theory of Semantic Communication

Semantic communication is an emerging research area that has gained a wide range of attention recently. Despite this growing interest, there remains a notable absence of a comprehensive and widely-accepted framework for characterizing semantic communication. This paper introduces a new conceptualization of semantic communication and formulates two fundamental problems, which we term language exploitation and language design. Our contention is that the challenge of language design can be effectively situated within the broader framework of joint source-channel coding theory, underpinned by a comprehensive end-to-end distortion metric. To tackle the language exploitation problem, we put forth three approaches: semantic encoding, semantic decoding, and a synergistic combination of both in the form of combined semantic encoding and decoding. Furthermore, we establish the semantic distortion-cost region as a critical framework for assessing the language exploitation problem. For each of the three proposed approaches, the achievable distortion-cost region is characterized. Overall, this paper aims to shed light on the intricate dynamics of semantic communication, paving the way for a deeper understanding of this evolving field.

cs.IT↗

arXiv2025

Disentangling the Drivers of LLM Social Conformity: An Uncertainty-Moderated Dual-Process Mechanism

As large language models (LLMs) integrate into collaborative teams, their social conformity -- the tendency to align with majority opinions -- has emerged as a key concern. In humans, conformity arises from informational influence (rational use of group cues for accuracy) or normative influence (social pressure for approval), with uncertainty moderating this balance by shifting from purely analytical to heuristic processing. It remains unclear whether these human psychological mechanisms apply to LLMs. This study adapts the information cascade paradigm from behavioral economics to quantitatively disentangle the two drivers to investigate the moderate effect. We evaluated nine leading LLMs across three decision-making scenarios (medical, legal, investment), manipulating information uncertainty (q = 0.667, 0.55, and 0.70, respectively). Our results indicate that informational influence underpins the models' behavior across all contexts, with accuracy and confidence consistently rising with stronger evidence. However, this foundational mechanism is dramatically modulated by uncertainty. In low-to-medium uncertainty scenarios, this informational process is expressed as a conservative strategy, where LLMs systematically underweight all evidence sources. In contrast, high uncertainty triggers a critical shift: while still processing information, the models additionally exhibit a normative-like amplification, causing them to overweight public signals (beta > 1.55 vs. private beta = 0.81).

cs.CY↗

arXiv2025

From Generation to Consumption: Personalized List Value Estimation for Re-ranking

Re-ranking is critical in recommender systems for optimizing the order of recommendation lists, thus improving user satisfaction and platform revenue. Most existing methods follow a generator-evaluator paradigm, where the evaluator estimates the overall value of each candidate list. However, they often ignore the fact that users may exit before consuming the full list, leading to a mismatch between estimated generation value and actual consumption value. To bridge this gap, we propose CAVE, a personalized Consumption-Aware list Value Estimation framework. CAVE formulates the list value as the expectation over sub-list values, weighted by user-specific exit probabilities at each position. The exit probability is decomposed into an interest-driven component and a stochastic component, the latter modeled via a Weibull distribution to capture random external factors such as fatigue. By jointly modeling sub-list values and user exit behavior, CAVE yields a more faithful estimate of actual list consumption value. We further contribute three large-scale real-world list-wise benchmarks from the Kuaishou platform, varying in size and user activity patterns. Extensive experiments on these benchmarks, two Amazon datasets, and online A/B testing on Kuaishou show that CAVE consistently outperforms strong baselines, highlighting the benefit of explicitly modeling user exits in re-ranking.

cs.IR↗

arXiv2025

Which Programming Language and What Features at Pre-training Stage Affect Downstream Logical Inference Performance?

Recent large language models (LLMs) have demonstrated remarkable generalization abilities in mathematics and logical reasoning tasks. Prior research indicates that LLMs pre-trained with programming language data exhibit high mathematical and reasoning abilities; however, this causal relationship has not been rigorously tested. Our research aims to verify which programming languages and features during pre-training affect logical inference performance. Specifically, we pre-trained decoder-based language models from scratch using datasets from ten programming languages (e.g., Python, C, Java) and three natural language datasets (Wikipedia, Fineweb, C4) under identical conditions. Thereafter, we evaluated the trained models in a few-shot in-context learning setting on logical reasoning tasks: FLD and bAbi, which do not require commonsense or world knowledge. The results demonstrate that nearly all models trained with programming languages consistently outperform those trained with natural languages, indicating that programming languages contain factors that elicit logic inference performance. In addition, we found that models trained with programming languages exhibit a better ability to follow instructions compared to those trained with natural languages. Further analysis reveals that the depth of Abstract Syntax Trees representing parsed results of programs also affects logical reasoning performance. These findings will offer insights into the essential elements of pre-training for acquiring the foundational abilities of LLMs.

cs.CL↗