Search arXiv⌕ Search

arXiv subjects

Artificial intelligence

Search indexed arXiv papers on artificial intelligence and machine learning, including cs.AI metadata. Follow the original manuscripts for methods, experiments and version history.

Follow this topic with RSS
At least 325 records · Page 18Linked to original sources
arXiv2026

A game theory for foundation models shows new paths to rational cooperation through similarity inference

As autonomous agents powered by foundation models are increasingly integrated into social and economic systems, understanding the principles governing their collective behavior is essential for ensuring safety and cooperation. Classical game theory, the dominant framework for modeling rational interaction, is built upon the assumption of `decoupled agency,' where agents treat their own decision-making as independent of the environment and other actors. Modern AI agents, however, jointly predict their own future actions alongside external observations. Here, we report a striking finding: when interacting in stylized social dilemmas, foundation model agents engaging in optimal planning consistently converge to stable cooperation, directly contradicting classical game-theoretic predictions of mutual defection. To understand this phenomenon, we introduce the `embedded Bayesian agent,' a theoretical model for foundation model agents. By shifting from decoupled to embedded agency, these agents model themselves as part of the universe they inhabit, maintaining epistemic uncertainty about their own decision-making algorithms. We show that by inferring whether others are behaviorally similar, an embedded agent treats its own deliberation during planning as evidence: a decision to cooperate predicts a similar decision by a similar partner. We formalize this mechanism of similarity inference through the `embedded equilibrium,' a novel solution concept replacing the Nash equilibrium to provide a foundational game theory for the social behavior of modern AI agents.

cs.AI↗
arXiv2026

Interpretable Adaptive Sampling for LLM Test-Time Scaling

Test-time scaling improves LLM reasoning by generating and aggregating multiple candidate answers, yet many pipelines use fixed per-query budgets that spend the same compute on easy and difficult prompts. These fixed budgets are also difficult to inspect because they do not explain why a given prompt receives a particular number of samples. We propose adaptive} test-time scaling with a lightweight fuzzy controller that maps interpretable signals, including estimated prompt complexity and model confidence, to a per-query sampling budget. The controller assigns fewer samples to easier or more confident prompts and more samples to harder or less certain prompts, making inference-time compute inspectable rather than fixed or opaque. We evaluate under a fair-alignment protocol with matched decoding settings and controlled answer selection, and compare against best-of-$N$, compute-aware scaling, and self-certainty-based baselines on question-answering and mathematical reasoning tasks. Across models and datasets, adaptive fuzzy control improves over several standard baselines and remains close to a selector-matched full-budget control while reducing the average number of samples. These findings suggest that interpretable adaptive sampling is a practical direction for more efficient test-time reasoning in large language models.

cs.AI↗
arXiv2026

Separating quantum circuits from classical LLMs

Modern large language models - transformers and diffusion language models - are built around two canonical algorithmic tasks: prediction and generation. We prove unconditional separations between low-depth quantum computation and the corresponding bounded-resource classical language-model architectures in both regimes. Concretely, we exhibit the following: 1. Distributional separation. We give a distribution that is sampleable by $\textsf{QNC}^0$ circuits (i.e., a family of constant-depth quantum circuits consisting of bounded fan-in gates) that no constant-round diffusion language model ($\textsf{DLM}$) with shallow scheduling and denoising can sample within constant distance, even when allowed sublinear chain-of-thought and output-token revision/remasking events, the very features modern $\textsf{DLM}$s rely on. 2. Functional separation. We exhibit a function computable in $\land \circ \textsf{QNC}^0[\log\log n]$ (i.e., a family of O$(\log\log n)$-depth $\textsf{QNC}^0$ circuits, where $n$ is the input length, followed by a single classical $\mathsf{AND}$ gate) such that any constant-depth decoder-only transformer computing the function must be large: it would have to have width $n^{Ω(1)}$. Together, our work initiates the study of quantum advantage in the era of large language models.

quant-ph↗
arXiv2026

Should We Type or Talk to LLM Agents? A Comprehensive Study of Voice and Keyboard Input Perturbations

Human input reaches language models by typing or speaking, and each channel leaves a distinct signature: orthographic noise for keyboards; for voice, disfluency from conventional transcription and restructuring from AI-backed dictation tools. How do they impact an LLM's performance? In this paper we present HIVE (Human Input-Variation Engine), a suite of voice transcription perturbations and QWERTY keyboard perturbations. We use HIVE to evaluate how robust models are to these perturbations. We present seven findings. (i) Voice transcription perturbations lower accuracy across every instruction-tuned model we test, and it is the structure of the transcription rather than its fillers that carries the cost. (ii) QWERTY keyboard perturbations cost less, and a model absorbs a lot of them before accuracy falls away. (iii) Both trace back to one cause, how many of the question's tokens survive the perturbation: destroying a token is what hurts, while adding new ones alongside it costs little. (iv) The gap between the two channels appears only where the answer must be constructed or deduced; on multiple choice there is none. (v) The harm does not solely come from test-set contamination. (vi) It cannot be trained away with lightweight adaptation. (vii) A thinking budget recovers the keyboard channel almost entirely but leaves the spoken registers untouched, and compressed speech is worse with it.

cs.AI↗
arXiv2026

ReflectRL: Learning from Golden Negative Trajectories via Reflective-to-Direct Reasoning

On-policy training has emerged as a powerful post-training paradigm for improving the reasoning capabilities of large language models, and is often enhanced by golden trajectories from stronger expert models. However, when the expert fails on harder problems, existing trajectory-guided methods lose their main source of supervision, and these failed trajectories are typically discarded as negative samples. We argue that such failures, which we call Golden Negative Trajectories, can still provide valuable reasoning signals when treated not as demonstrations to imitate, but as flawed trajectories to reflect upon. We identify a Reflection Advantage: for hard problems, reflecting on a flawed trajectory can be easier and more effective than solving the problem directly from scratch. Motivated by this, we propose ReflectRL, a lightweight plug-and-play framework that learns from Golden Negative Trajectories during on-policy training. ReflectRL first uses these trajectories to elicit Reflective Reasoning, then applies Reflective-to-Direct Policy Transition to transfer the acquired reasoning behavior back to Direct Reasoning. Experiments across 9 benchmarks, 4 LLM backbones, and 4 on-policy training methods show that ReflectRL consistently improves reasoning performance with minimal overhead.

cs.AI↗
arXiv2026

Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent

We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. Preliminary evaluations reveal two critical bottlenecks in current models: (1) modality bias, where agents bypass visual tools in favor of textual search, and (2) parametric knowledge leakage, where models rely on internal memory rather than genuine tool-augmented execution. To address these challenges, we propose Video-DR, featuring a decoupled perception-exploration pipeline with stage-wise tool unlocking that compels exhaustive cross-frame visual grounding prior to web retrieval. Our framework adopts a two-stage training recipe: supervised fine-tuning followed by Group Relative Policy Optimization (GRPO), enabling autonomous exploration that breaks the imitation-learning ceiling. Furthermore, we curate Video-DR-Bench, a human-AI collaborative benchmark comprising 200 complex, multi-hop VQA instances. Empirical results demonstrate that our Video-DeepResearch-35B-A3B establishes a new state-of-the-art of 64.0% average accuracy, surpassing proprietary Claude-4.5-Sonnet (59.0%) by 5.0 points and significantly outperforming GPT-5 (52.5%) and Gemini 2.5 Pro (57.5%). The 30B-A3B variant achieves 59.3%, competitive with Claude-4.5-Sonnet and demonstrating the effectiveness of our training paradigm even at compact scale. Code: https://github.com/Osilly/Vision-DeepResearch.

cs.CV↗
arXiv2026

TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning

Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, limiting fine-grained credit assignment in long-horizon TIR scenarios. On-policy self-distillation offers denser signals through teacher branches with privileged context, but existing approaches typically derive such context from ground-truth answers or retrieved skills, which may not reflect the states actually visited by the agent. Moreover, token-level supervision fails to capture the turn-level structure of tool interactions. To address this, we propose TurnSight, a turn-level hindsight self-distillation framework that derives supervision directly from execution-conditioned hindsight. It then constructs multiple hindsight views with different lookahead horizons and selects reliable supervision through cross-horizon directional agreement. Finally, the selected hindsight signal is normalized across sibling rollouts and used to adaptively modulate RL advantages while preserving their original optimization direction. Extensive experiments on three benchmarks demonstrate the effectiveness of TurnSight. Our codes are available at https://github.com/quchangle1/TurnSight.

cs.CL↗
arXiv2026

Escaping Python Dependency Hell: A Hybrid Replay-and-Repair Pipeline for Python Dependency Resolution

Dependency conflicts in Python ecosystems arise from incompatible version constraints, missing packages, and undocumented compatibility relationships, causing many real-world code snippets to fail at execution. This paper presents PLLM+, a hybrid dependency-repair pipeline evaluated on the HG2.9K benchmark of 2,891 dependency-failing snippets. PLLM+ prioritizes inexpensive deterministic steps before invoking LLM-based repair: static AST-based interpreter inference, replay of historically successful dependency configurations from the competition-provided solutions database, and live PyPI validation of candidate package versions. When these steps do not resolve a case, the system falls back to a structured LLM-based repair loop with typed error classification and Proposer/Critic agents. On HG2.9K, PLLM+ solves 1,500 out of 2,891 snippets, compared with 1,169 solved by the PLLM baseline. It also reduces average runtime from 368.7 to 71.8 seconds per snippet. Most successful fixes come from replaying known configurations: 1,495 of the 1,500 successful fixes are produced by the solutions database, while the LLM fallback accounts for 5 additional fixes. These results suggest that, in this benchmark setting, deterministic reuse of previously validated dependency configurations is a simple and effective strategy, with LLM-based repair serving as a secondary fallback for cases not covered by prior solutions.

cs.AI↗
arXiv2026

Security in the Fine-Tuning Lifecycle of Large Language Models: Threats, Defenses,Evaluation, and Future Directions

Background: Fine-tuning is central to adapting pre-trained Large Language Models (LLMs) to downstream tasks, but its reliance on training data, parameter updates, and reusable components opens entry points for attackers. Threats have evolved from data poisoning and weight tampering to agent manipulation and interface exploitation, yet existing reviews lack a unified framework spanning the full fine-tuning lifecycle. Objective: This paper presents a systematic survey of LLM fine-tuning security and establishes a lifecycle-based framework for comparing attacks and defenses, complemented by unified empirical evaluation. Methods: We divide attack and defense mechanisms into three phases by intervention timing: pre-tuning, during-tuning, and post-tuning. Within each phase, strategies are reviewed and contrasted to expose their evolution and limitations. Representative methods are then evaluated under a unified model, hardware, and protocol setup, with cross-phase experiments pairing attacks and defenses from different phases. Results: Attack effectiveness is highly model-dependent and non-monotonic with scale: weight-editing attacks effective on earlier models lose impact on modern open-source LLMs; cross-lingual backdoor transfer, reported as near-perfect at larger scales, fails entirely on tested 1B-4B models; and purely benign samples can compromise safety alignment in instruction-tuned models. Single-phase defenses rarely generalize across phases, and defense effectiveness depends jointly on model architecture and alignment state. Conclusion: We identify key open problems (configuration-robust defense, cross-phase defense composition, and embedding-space attacks beyond behavioral assumptions) and propose concrete future research directions.

cs.CR↗
arXiv2026

Automated Visualization Code Synthesis via Multi-Path Reasoning and Feedback-Driven Optimization

Large Language Models (LLMs) have become a cornerstone for automated visualization code generation, enabling users to create charts through natural language instructions. Despite improvements from techniques like few-shot prompting and query expansion, existing methods often struggle when requests are underspecified in actionable details (e.g., data preprocessing assumptions, solver or library choices, etc.), frequently necessitating manual intervention. To overcome these limitations, we propose VisPath: a Multi-Path Reasoning and Feedback-Driven Optimization Framework for Visualization Code Generation. VisPath handles underspecified queries through structured, multi-stage processing. It begins by using Chain-of-Thought (CoT) prompting to reformulate the initial user input, generating multiple extended queries in parallel to surface alternative plausible concretizations of the request. These queries then generate candidate visualization scripts, which are executed to produce diverse images. By assessing the visual quality and correctness of each output, VisPath generates targeted feedback that is aggregated to synthesize an optimal final result. Extensive experiments on MatPlotBench and Qwen-Agent Code Interpreter Benchmark show that VisPath outperforms state-of-the-art methods, providing a more reliable framework for AI-driven visualization generation.

cs.SE↗
arXiv2025

Overview of CHIP 2025 Shared Task 2: Discharge Medication Recommendation for Metabolic Diseases Based on Chinese Electronic Health Records

Discharge medication recommendation plays a critical role in ensuring treatment continuity, preventing readmission, and improving long-term management for patients with chronic metabolic diseases. This paper present an overview of the CHIP 2025 Shared Task 2 competition, which aimed to develop state-of-the-art approaches for automatically recommending appro-priate discharge medications using real-world Chinese EHR data. For this task, we constructed CDrugRed, a high-quality dataset consisting of 5,894 de-identified hospitalization records from 3,190 patients in China. This task is challenging due to multi-label nature of medication recommendation, het-erogeneous clinical text, and patient-specific variability in treatment plans. A total of 526 teams registered, with 167 and 95 teams submitting valid results to the Phase A and Phase B leaderboards, respectively. The top-performing team achieved the highest overall performance on the final test set, with a Jaccard score of 0.5102, F1 score of 0.6267, demonstrating the potential of advanced large language model (LLM)-based ensemble systems. These re-sults highlight both the promise and remaining challenges of applying LLMs to medication recommendation in Chinese EHRs. The post-evaluation phase remains open at https://tianchi.aliyun.com/competition/entrance/532411/.

cs.CL↗
arXiv2026

Curiosity-Diffuser: Curiosity Guide Diffusion Models for Reliability

One of the bottlenecks in robotic intelligence is the instability of neural network models. This leads to risks when applying intelligence in the physical world. Specifically, imitation policy based on neural network may generate hallucinations, leading to inaccurate behaviors that impact the safety of real-world applications. To address this issue, this paper proposes the Curiosity-Diffuser, aimed at guiding the conditional diffusion model to generate trajectories with lower curiosity, thereby improving the reliability of policy. The core idea is to use a Random Network Distillation (RND) curiosity module to assess whether the model's behavior aligns with the training data, and then minimize curiosity by classifier guidance diffusion to reduce overgeneralization during inference. Additionally, we propose a computationally efficient metric for evaluating the reliability of the policy, measuring the similarity between the generated behaviors and the training dataset, to facilitate research about reliability learning. Finally, simulations and real-world experiments verify the effectiveness and applicability of the proposed method to a variety of scenarios, showing that Curiosity-Diffuser significantly improves task performance and produces behaviors that are more similar to the training data. The code for this work is available at: github.com/CarlDegio/Curiosity-Diffuser

cs.RO↗
arXiv2026

Efficient unsupervised domain adaptation via self-supervised vision transformer and synergistic cross-domain alignment

Unsupervised domain adaptation (UDA) aims to mitigate domain shift, where the distribution of labeled source data differs from that of unlabeled target data. Despite recent advances, existing methods often rely on fine-tuning large backbone models, which leads to high computational cost and limits scalability in resource-constrained environments. This limitation highlights the need for parameter-efficient approaches that maintain strong performance with reduced training complexity. Self-supervised foundation models such as DINOv2 provide highly transferable representations and raise the question of whether effective domain adaptation can be achieved without full fine-tuning. To address this question, we propose Efficient Unsupervised Domain Adaptation (EUDA), a parameter-efficient framework that leverages a frozen DINOv2 backbone as a feature extractor and updates only a lightweight bottleneck and classification head. We also adopt a synergistic domain alignment loss (SDAL), which combines cross-entropy (CE) and maximum mean discrepancy (MMD) to promote both discriminative learning and cross-domain alignment. Experimental results on Office-Home, Office-31, VisDA-2017, and DomainNet demonstrate that EUDA achieves competitive performance across diverse domain complexities, while reducing the number of trainable parameters by 42 to 99.7%. These results show the suitability of the proposed method for resource-constrained and distributed environments.

cs.CV↗
arXiv2025

ArchiveGPT: A human-centered evaluation of using a vision language model for image cataloguing

The accelerating growth of photographic collections has outpaced manual cataloguing, motivating the use of vision language models (VLMs) to automate metadata generation. This study examines whether Al-generated catalogue descriptions can approximate human-written quality and how generative Al might integrate into cataloguing workflows in archival and museum collections. A VLM (InternVL2) generated catalogue descriptions for photographic prints on labelled cardboard mounts with archaeological content, evaluated by archive and archaeology experts and non-experts in a human-centered, experimental framework. Participants classified descriptions as AI-generated or expert-written, rated quality, and reported willingness to use and trust in AI tools. Classification performance was above chance level, with both groups underestimating their ability to detect Al-generated descriptions. OCR errors and hallucinations limited perceived quality, yet descriptions rated higher in accuracy and usefulness were harder to classify, suggesting that human review is necessary to ensure the accuracy and quality of catalogue descriptions generated by the out-of-the-box model, particularly in specialized domains like archaeological cataloguing. Experts showed lower willingness to adopt AI tools, emphasizing concerns on preservation responsibility over technical performance. These findings advocate for a collaborative approach where AI supports draft generation but remains subordinate to human verification, ensuring alignment with curatorial values (e.g., provenance, transparency). The successful integration of this approach depends not only on technical advancements, such as domain-specific fine-tuning, but even more on establishing trust among professionals, which could both be fostered through a transparent and explainable AI pipeline.

cs.HC↗
arXiv2026

Compass: Optimizing Compound AI Workflows for Dynamic Adaptation

Compound AI is a distributed intelligence approach that represents a unified system orchestrating specialized AI/ML models with engineered software components into AI workflows. Compound AI production deployments must satisfy accuracy, latency, and cost objectives under varying loads. However, many deployments operate on fixed infrastructure where horizontal scaling is not viable. Existing approaches optimize solely for accuracy and do not consider changes in workload conditions. We observe that compound AI systems can switch between configurations to fit infrastructure capacity, trading accuracy for latency based on current load. This requires discovering multiple Pareto-optimal configurations from a combinatorial search space and determining when to switch between them at runtime. We present Compass, a novel framework that enables dynamic configuration switching through offline optimization and online adaptation. Compass consists of three components: COMPASS-V algorithm for configuration discovery, Planner for switching policy derivation, and Elastico Controller for runtime adaptation. COMPASS-V discovers accuracy-feasible configurations using finite-difference guided search and a combination of hill-climbing and lateral expansion. Planner profiles these configurations on target hardware and derives switching policies using a queuing theory based model. Elastico monitors queue depth and switches configurations based on derived thresholds. Across two compound AI workflows, COMPASS-V achieves 100% recall while reducing configuration evaluations by 57.5% on average compared to exhaustive search, with efficiency gains reaching 95.3% at tight accuracy thresholds. Runtime adaptation achieves 90-98% SLO compliance under dynamic load patterns, improving SLO compliance by 71.6% over static high-accuracy baselines, while simultaneously improving accuracy by 3-5% over static fast baselines.

cs.DC↗
arXiv2024

Crowdsourced Multilingual Speech Intelligibility Testing

With the advent of generative audio features, there is an increasing need for rapid evaluation of their impact on speech intelligibility. Beyond the existing laboratory measures, which are expensive and do not scale well, there has been comparatively little work on crowdsourced assessment of intelligibility. Standards and recommendations are yet to be defined, and publicly available multilingual test materials are lacking. In response to this challenge, we propose an approach for a crowdsourced intelligibility assessment. We detail the test design, the collection and public release of the multilingual speech data, and the results of our early experiments.

eess.AS↗
arXiv2021

Verdi: Quality Estimation and Error Detection for Bilingual Corpora

Translation Quality Estimation is critical to reducing post-editing efforts in machine translation and to cross-lingual corpus cleaning. As a research problem, quality estimation (QE) aims to directly estimate the quality of translation in a given pair of source and target sentences, and highlight the words that need corrections, without referencing to golden translations. In this paper, we propose Verdi, a novel framework for word-level and sentence-level post-editing effort estimation for bilingual corpora. Verdi adopts two word predictors to enable diverse features to be extracted from a pair of sentences for subsequent quality estimation, including a transformer-based neural machine translation (NMT) model and a pre-trained cross-lingual language model (XLM). We exploit the symmetric nature of bilingual corpora and apply model-level dual learning in the NMT predictor, which handles a primal task and a dual task simultaneously with weight sharing, leading to stronger context prediction ability than single-direction NMT models. By taking advantage of the dual learning scheme, we further design a novel feature to directly encode the translated target information without relying on the source context. Extensive experiments conducted on WMT20 QE tasks demonstrate that our method beats the winner of the competition and outperforms other baseline methods by a great margin. We further use the sentence-level scores provided by Verdi to clean a parallel corpus and observe benefits on both model performance and training efficiency.

cs.CL↗
arXiv2021

Zero Shot on the Cold-Start Problem: Model-Agnostic Interest Learning for Recommender Systems

User behavior has been validated to be effective in revealing personalized preferences for commercial recommendations. However, few user-item interactions can be collected for new users, which results in a null space for their interests, i.e., the cold-start dilemma. In this paper, a two-tower framework, namely, the model-agnostic interest learning (MAIL) framework, is proposed to address the cold-start recommendation (CSR) problem for recommender systems. In MAIL, one unique tower is constructed to tackle the CSR from a zero-shot view, and the other tower focuses on the general ranking task. Specifically, the zero-shot tower first performs cross-modal reconstruction with dual auto-encoders to obtain virtual behavior data from highly aligned hidden features for new users; and the ranking tower can then output recommendations for users based on the completed data by the zero-shot tower. Practically, the ranking tower in MAIL is model-agnostic and can be implemented with any embedding-based deep models. Based on the co-training of the two towers, the MAIL presents an end-to-end method for recommender systems that shows an incremental performance improvement. The proposed method has been successfully deployed on the live recommendation system of NetEase Cloud Music to achieve a click-through rate improvement of 13% to 15% for millions of users. Offline experiments on real-world datasets also show its superior performance in CSR. Our code is available.

cs.IR↗