Academic Intelligence · Curated Daily

探索全球前沿学术脉络

AcademicHub 汇聚顶级期刊与预印本平台的实时文献。定制您的专属科研雷达,利用大语言模型自动生成交叉领域文献分析简报。

01.
arXiv (CS.CL) 2026-06-12

NOVA: NOise-aware Verbal Confidence CAlibration for Robust Large Language Models in RAG Systems

Accurately assessing model confidence is essential for deploying large language models (LLMs) in mission-critical factual domains. While retrieval-augmented generation (RAG) is widely adopted to improve grounding, confidence calibration in RAG settings remains poorly understood. We conduct a systematic study across four benchmarks, revealing that LLMs exhibit poor calibration performance especially when noisy contexts are retrieved. Specifically, contradictory or irrelevant evidence tends to exacerbate the model's overconfidence issue. To address this, we propose NOVA Rules (NOise-Aware Verbal Confidence CAlibration Rules) to provide a principled foundation for resolving overconfidence under noise. We further design NOVA, a noise-aware calibration framework that synthesizes supervision from ~2K HotpotQA examples guided by these rules. By performing supervised fine-tuning (SFT) with this data, NOVA equips models with intrinsic noise awareness without relying on stronger teacher models. Empirical results show that NOVA yields substantial gains, improving ECE scores by 10.9% in-domain and 8.0% out-of-domain. By bridging the gap between retrieval noise and verbal calibration, NOVA paves the way for both accurate and epistemically reliable LLMs.

02.
arXiv (CS.LG) 2026-06-16

Constraining the outputs of ReLU neural networks

arXiv:2508.03867v2 Announce Type: replace-cross Abstract: We introduce a class of algebraic varieties naturally associated with ReLU neural networks, arising from the piecewise linear structure of their outputs across activation regions in input space, and the piecewise multilinear structure in parameter space. By analyzing the rank constraints on the network outputs within each activation region, we derive polynomial equations that characterize the functions representable by the network. We further investigate conditions under which these varieties attain their expected dimension, providing insight into the expressive and structural properties of ReLU networks.

04.
arXiv (CS.CL) 2026-06-18

Evaluating Prompting-Based Defenses Against Domain-Camouflaged Injection Attacks

作者:

Domain-camouflaged injection attacks embed malicious instructions in retrieved content using domain-appropriate vocabulary, evading standard detectors that rely on syntactic injection markers. When detection fails, practitioners need to know which defense architectures reduce attack success. We evaluate five prompting-based defenses (spotlighting, paraphrasing, prompt sandwiching, and two combinations) against domain-camouflaged injection across three model families (Claude Haiku, Llama 3.1 8B, Gemini 2.0 Flash) and three deployment domains (financial, legal, general) using 3,510 trials. Paraphrasing retrieved content before agent processing is the most consistently effective defense in this benchmark, reducing camouflage attack success rate by 55-84\% depending on model, and achieves lower attack success rates than our Llama Guard 4 configuration on every model tested. Defense effectiveness is strongly model-dependent: spotlighting halves attack success on Claude Haiku but provides no benefit on Llama 3.1 8B. Financial domain deployments face the highest residual risk at 26-33\% baseline attack success rate, with no prompting-based defense fully eliminating the threat on weaker models. These results provide the first systematic evaluation of prompting-based defenses specifically against camouflage-class injection attacks and establish benchmark-based recommendations for practitioners. All tasks use synthetically constructed professional documents; whether these benchmark rankings generalize to real enterprise documents remains an open question.

05.
arXiv (CS.CL) 2026-06-16

QK-Normed MLA: QK normalization without full key caching

Query-key (QK) normalization stabilizes attention by controlling the scale of queries and keys before the dot product, but is not immediately compatible with Multi-head Latent Attention (MLA). MLA achieves efficient decoding by caching low-dimensional latent states instead of full keys, whereas post-projection QK RMSNorm appears to require the fully projected key for every cached token. We show this apparent incompatibility is an implementation artifact, not an architectural constraint. RMSNorm decomposes into a static affine weight and a dynamic scalar RMS statistic. The static key-side weight can be absorbed into the MLA query-side projection; the dynamic key statistic reduces to one inverse-RMS scalar per token and KV group. The resulting formulation is exactly equivalent to explicit post-projection QK RMSNorm in exact arithmetic and preserves MLA's latent decode path. In our 400M runs trained for up to 100B tokens, QK-Normed MLA achieves lower training loss and better downstream accuracy than QK clipping, while H800 decode benchmarks show less than 2% latency overhead up to 256k context. These results make QK normalization a practical stabilization option for MLA models without requiring full-key caching.

06.
arXiv (CS.LG) 2026-06-11

Prediction-Powered Risk Monitoring of Deployed Models for Detecting Harmful Distribution Shifts

arXiv:2602.02229v2 Announce Type: replace Abstract: We study the problem of monitoring model performance in dynamic environments where labeled data are limited. To this end, we propose prediction-powered risk monitoring (PPRM), a semi-supervised risk-monitoring approach based on prediction-powered inference (PPI). PPRM constructs anytime-valid lower bounds on the running risk by combining synthetic labels with a small set of true labels. Harmful shifts are detected via a threshold-based comparison with an upper bound on the nominal risk, satisfying assumption-free finite-sample guarantees on the type-I error. We demonstrate the effectiveness of PPRM through extensive experiments on image classification, large language model (LLM), and telecommunications monitoring tasks.

07.
arXiv (CS.AI) 2026-06-16

MUZZLE: Adaptive Agentic Red-Teaming of Web Agents Against Indirect Prompt Injection Attacks

arXiv:2602.09222v2 Announce Type: replace-cross Abstract: Large language model (LLM) based web agents are increasingly deployed to automate complex online tasks by directly interacting with web sites and performing actions on users' behalf. While these agents offer powerful capabilities, their design exposes them to indirect prompt injection attacks embedded in untrusted web content, enabling adversaries to hijack agent behavior and violate user intent. Despite growing awareness of this threat, existing evaluations rely on fixed attack templates, manually selected injection surfaces, or narrowly scoped scenarios, limiting their ability to capture realistic, adaptive attacks encountered in practice. We present MUZZLE, an automated agentic framework for evaluating the security of web agents against indirect prompt injection attacks. MUZZLE utilizes the agent's trajectories to automatically identify high-salience injection surfaces, and adaptively generate context-aware malicious instructions that target violations of confidentiality, integrity, and availability. Unlike prior approaches, MUZZLE adapts its attack strategy based on the agent's observed execution trajectory and iteratively refines attacks using feedback from failed executions. We evaluate MUZZLE across diverse web applications, user tasks, and agent configurations, demonstrating its ability to automatically and adaptively assess the security of web agents with minimal human intervention. Our results show that MUZZLE effectively discovers 44 new attacks on 4 web applications with 10 adversarial objectives that violate confidentiality, availability, or privacy properties across different LLMs and agent scaffolds. MUZZLE also identifies novel attack strategies, including 3 cross-application prompt injection attacks and an agent-tailored phishing scenario.

08.
arXiv (CS.CL) 2026-06-19

Efficiently Representing Algorithms With Chain-of-Thought Transformers

The increasing popularity of reasoning models – language models that output a series of reasoning or thought tokens before producing an answer – is justified, in part, by theoretical results showing that chain-of-thought (CoT) transformers can simulate Turing machines, and thus perform arbitrary computation. However, the Turing machine, while suitable for complexity-theoretic analysis, is not convenient, intuitive, or efficient for discussing algorithms. Algorithms are typically designed and analyzed at a higher level of abstraction, captured by the Word RAM model with random-access memory and unit-cost operations on $\bigO(\log n)$-bit words. As a result, Word RAM algorithms can be substantially more efficient than their Turing machine counterparts, raising the question: Can CoT transformers efficiently simulate Word RAM algorithms? For instance, can they sort $n$ items in $\bigO(n \log n)$ steps or run Dijkstra's algorithm in $\bigO(E + V \log V)$ steps? We answer affirmatively, up to poly-logarithmic overhead. We first establish this for finite-precision transformers with poly-logarithmic width and rightmost unique hard attention, then strengthen the result to two more practical settings with finite width and log-precision: continuous CoT, where reasoning takes the form of vectors rather than tokens, and a hybrid architecture in which transformer layers sit atop a recurrent (linear RNN) layer. In all three cases, we find that CoT can efficiently simulate any Word RAM algorithm with only a poly-logarithmic overhead in $n$. This overhead reduces to log-square when the Word RAM has a ``flat'' instruction set, and only logarithmic for multiplication-free flat instructions – in stark contrast to known CoT simulations of Turing machines, which require quadratic overhead over Word RAM.

09.
arXiv (CS.CV) 2026-06-11

Adv-TGD: Adversarial Text-Guided Diffusion for Face Recognition Impersonation Attacks

The widespread adoption of face recognition (FR) technologies raises serious privacy concerns, as facial data can be exploited without consent. To address this challenge, we propose Adv-TGD, a generative adversarial attack framework that synthesizes photorealistic faces capable of impersonating target identities and deceiving face recognition systems. Built upon Stable Diffusion, Adv-TGD performs per-sample LoRA fine-tuning conditioned on concise textual prompts to generate natural yet adversarially manipulated identities. Unlike conventional identity-attack approaches, our method optimizes lightweight cross-attention adapters for each source-target pair within a single-step denoising process. Latent blending is constrained by a face-local heatmap mask to ensure spatially precise identity manipulation while preserving non-sensitive regions. We introduce a composite objective that integrates masked epsilon-MSE reconstruction, thresholded identity divergence in FR embedding space, directional feature alignment, and source-similarity suppression to balance adversarial attack and visual realism. Optionally, LLaVA-generated attribute prompts enhance fine-grained semantic details without reintroducing identity cues. Under the black-box evaluation protocol, Adv-TGD attains an average attack success rate (ASR) of 85.90% across IR152, IRSE50, MobileFace, and FaceNet, surpassing the semantic SOTA baseline Adv-CPG by +6.25 points, diffusion-based makeup method DiffAIM by +3 points, and noise-based P3-Mask by +16 points. Despite its strong attack efficacy, Adv-TGD preserves high visual fidelity (PSNR = 27.15 dB, SSIM = 0.981). Furthermore, we demonstrate the flexibility of our framework by successfully extending it to in-the-wild datasets (LADN), general object classification (ImageNet), and transformer-based diffusion models (FLUX.1).

11.
PLOS Medicine 2026-05-20

Prescribed hormonal contraceptive use trends in the Estonian Biobank: A longitudinal observational study

by Jelisaveta Džigurski, Märt Möls, Kristi Läll, Hannah Currant, Mall Eltermaa, Estonian Biobank Research Team , Reedik Mägi, Lili Milani, Triin Laisk Background Hormonal contraceptives (HCs) are widely used and have well-documented population-level statistics. Previous studies with short follow-ups have focussed on individual HC use and side effects. However, the same aspects over longer periods, HC formulation switching, and the impact of genetic factors on HC side effects remain understudied due to the limited availability of suitable datasets. We investigated whether the Estonian Biobank (EstBB) is suitable for studying genetic risk for HC side effects. Methods and findings This is a longitudinal descriptive study combining prescribed HC purchase data collected from 2004 to 2022 with genetic and health data from 73,071 female EstBB HC users aged 15–55 at the time of purchase. HC usage was defined by the Anatomical Therapeutic Chemical (ATC) codes G02B, G03A, and G03HB01. Methods included calculating age-stratified annual user prevalence, inferring usage periods from purchases, assessing formulation switching, identifying the International Classification of Diseases, Tenth Revision (ICD-10)-based side effect-related diagnoses and thromboembolism risk factors, and assessing carrier status for Factor V Leiden (FVL, rs6025) and prothrombin G20210A (PTM, rs1799963) genetic variants as proof-of-concept. Over 19 years, 20 HC formulations with five administration routes (oral pills, transdermal patches, vaginal rings, subdermal implants, intrauterine devices) were used. In the EstBB, combined HCs were the most commonly used among users aged 15–29, while progestin-only HC use increased with age and over time, comparable to the Estonian population. Overall, 64.2% (n = 46,920) of users switched formulations at least once, with 17.7% (n = 12,929) being rapid switchers. Side effect-related diagnoses were observed in 23.1% (n = 2,982) of rapid switchers, with excessive/irregular menstrual bleeding being the most common. Genetic analysis revealed that 5.3% (n = 3,886) of users carried at least one variant previously associated with increased thrombosis risk (3.5% (n = 2,556) carried FVL only, 1.8% (n = 1,276) PTM only, and 0.07% (n = 54) both). Carriers of thrombosis-associated variants had a significantly higher percentage of thrombosis (6.5%) than non-carriers (4.2%; OR = 1.61, 95% CI [1.40, 1.84], p 

12.
arXiv (quant-ph) 2026-06-11

Robust Mixed-State Cluster States and Spurious Topological Entanglement Negativity

arXiv:2504.16165v2 Announce Type: replace Abstract: We investigate 1D and 2D cluster states under local decoherence to assess the robustness of their mixed-state subsystem symmetry-protected topological (SSPT) order. By exactly computing fidelity correlators via dimensional reduction of effective statistical mechanics models, we pinpoint the critical error rate for strong-to-weak spontaneous breaking of strong subsystem symmetry. Without resorting to the replica trick, we demonstrate that mixed-state SSPT order remains remarkably robust up to the maximal decoherence rate when noise respects strong subsystem symmetry. Furthermore, we propose that the mixed-state SSPT order can be detected by a constant correction to the area-law scaling of entanglement negativity, termed spurious topological entanglement negativity. This also highlights that topological entanglement negativity, a widely used diagnostic for mixed-state topological order, is generally not invariant under finite-depth quantum channels.

13.
arXiv (CS.LG) 2026-06-16

Formalizing and Mitigating Structural Distortion in LLM Attention for Zero-Shot Graph Reasoning

arXiv:2606.15633v1 Announce Type: new Abstract: Large Language Models (LLMs) have shown promise for reasoning over Text-Attributed Graphs (TAGs). However, applying LLMs to graphs requires linearizing their structure into sequences, introducing distortion rooted in the graph bandwidth problem. While this distortion has been shown to degrade performance, it is often attributed to prompt design or model scale, leaving the underlying mechanism unclear. In this work, we show how rotary positional embeddings turn graph linearization into bandwidth-dependent attention decay, suppressing attention between graph-adjacent nodes that are forced far apart in the serialized sequence. This shifts the focus of LLM-based graph reasoning from prompt engineering and scaling toward correcting attention misalignment. Motivated by this analysis, we propose Graph-aligned Language Attention (GaLA), a lightweight, inference-time modification for LLMs. GaLA biases attention toward graph-adjacent nodes while preserving the LLM's sequential inductive biases. Across TAG benchmarks, GaLA improves performance with negligible overhead, demonstrating that distortion is a correctable bottleneck in LLM-based graph reasoning.

14.
arXiv (CS.AI) 2026-06-12

Boosting Direct Preference Optimization with Penalization

作者:

arXiv:2606.12505v1 Announce Type: cross Abstract: Offline preference optimization has become a practical substitute for reinforcement learning from human feedback, but pairwise objectives such as Direct Preference Optimization (DPO) and its variants use only the chosen and rejected responses stored in a static dataset. This leaves a useful signal unused: the response that the reference model itself would generate for the same prompt. We propose Direct Preference Optimization with Penalization (DPOP), a simple extension of DPO that augments the base preference loss with a gated penalty on reference-greedy responses. DPOP activates this penalty only when the current policy still assigns a lower likelihood to the preferred response than to the rejected response. On AlpacaEval 2.0, DPOP improves length-controlled win rate over DPO, SimPO, and AlphaDPO on both Llama-3-8b-it and Gemma-2-9b-it, achieving relative gains of 5.3\% and 4.4\% over baselines on the two models, respectively. Ablations further show that a SimNPO-style length-normalized penalty is stronger than NPO and token-level unlikelihood in this setting.

15.
arXiv (CS.AI) 2026-06-16

Revisiting Chebyshev Polynomial and Anisotropic RBF Models for Tabular Regression

arXiv:2602.22422v2 Announce Type: replace-cross Abstract: Smooth-basis models such as Chebyshev polynomial regressors and radial basis function (RBF) networks are well established in numerical analysis. Their continuously differentiable prediction surfaces suit surrogate optimisation, sensitivity analysis, and other settings where the response varies gradually with inputs. Despite these properties, smooth models seldom appear in tabular regression, where tree ensembles dominate. We ask whether they can compete, benchmarking models across 55 regression datasets organised by application domain. We develop an anisotropic RBF network with data-driven centre placement and gradient-based width optimisation, a ridge-regularised Chebyshev polynomial regressor, and a smooth-tree hybrid (Chebyshev model tree); all three are released as scikit-learn-compatible packages. We benchmark these against tree ensembles, a pre-trained transformer, and standard baselines, evaluating accuracy alongside generalisation behaviour. The transformer ranks first on accuracy across a majority of datasets, but its GPU dependence, inference latency, and dataset-size limits constrain deployment in the CPU-based settings common across applied science and industry. Among CPU-viable models, smooth models and tree ensembles are statistically tied on accuracy, but the former tend to exhibit tighter generalisation gaps. We recommend routinely including smooth-basis models in the candidate pool, particularly when downstream use benefits from tighter generalisation and gradually varying predictions.

16.
arXiv (CS.AI) 2026-06-18

Closing the Loop: PID Feedback Control for Interpretable Activation Steering in Symbolic Music Generation

arXiv:2606.18790v1 Announce Type: cross Abstract: Transformer-based architectures have significantly advanced the generation of complex symbolic sequences, yet a significant gap remains in achieving fine-grained, interpretable control over discrete signal attributes. This paper investigates the mechanistic interpretability of the Multitrack Music Transformer (MMT) and proposes a framework for deterministic attribute modulation without retraining to bridge this gap via inference-time activation steering. Utilizing the Difference-in-Means (DiffMean) methodology, we isolate latent directions for signal attributes, specifically Pitch and Duration, within the residual stream. We validate the Linear Representation Hypothesis in this domain, achieving high correlation between steering magnitude and attribute shift. To address the inherent feature entanglement in multi-attribute steering, we introduce a Dual Steering framework utilizing Gram-Schmidt Orthogonalization. Experimental results demonstrate that this geometric decoupling reduces conceptual interference and signal degradation compared to naive vector addition, enabling independent deterministic control even against strong autoregressive conditioning.

17.
medRxiv (Medicine) 2026-06-18

The Effectiveness of aromatherapy and its supportive Interventions on anxiety and pain among breast cancer patients: A systematic review and meta-analysis

Introduction: Breast cancer treatments are often associated with pain and anxiety, which can hinder physical functioning and overall quality of life, even after treatment. Complementary therapies, such as aromatherapy, can be used to alleviate pain and reduce anxiety in breast cancer patients. This project aimed to synthesize current global evidence on the effectiveness of aromatherapy. Method: This systematic review followed the PRISMA 2020 guidelines, with a comprehensive, systematic search conducted in PubMed, CINAHL, Cochrane Library, and SCOPUS for randomized controlled trials (RCTS) published from 2015 to 2025. Eligible studies included adult women breast cancer surgery patients who received aromatherapy during various periods of breast cancer. Where possible, data from the included studies were pooled using meta-analysis. GRADE approach was used to assess certainty of findings. Results: The search yielded 84 studies. Out of these, six were included in this review. On average, aromatherapy reduces pain and anxiety scores by 0.79 (standard mean difference (SMD)=-0.79, 95% CI -1.42, -0.16) and 0.53 (SMD=-0.53, 95 CI=-0.90, -0.16) units, respectively, compared to control condition [Low-quality of evidence]. The combination of aromatherapy with music reduces pain and anxiety by 1.26 (SMD= -1.26, 95 CI=-1.65, -0.87) and 1.08 (SMD = -1.08, 95 % CI: -1.45, -0.70) units respectively compared to standard care [Low-quality of evidence]. Conclusion: There is a potential role for the use of aromatherapy and music therapy, to alleviate anxiety and pain, especially for non-preoperative anxiety and pain. Further research is needed to inform the integration of aromatherapy into the management of anxiety and pain.

18.
bioRxiv (Bioinfo) 2026-06-16

FlowBench: separating planning, fault recovery and interpretation in agentic bioinformatics

Agentic large language model (LLM) systems are being deployed in bioinformatics faster than they are understood, and single-metric evaluations conflate capabilities that fail independently. We introduce FlowBench, a benchmark that decomposes agentic bioinformatics performance into planning, fault recovery, biological interpretation, and end-to-end output-fidelity. Existing systems achieve high plan completeness, but their closed, single-provider designs prevent attribution of performance to scaffolding versus the underlying model. We therefore built FlowAgent, a modular, provider-agnostic framework whose components can be selectively disabled and whose backbone model can be swapped across providers on a shared harness, and used it to evaluate 23 models from three main providers. Three findings emerge. First, generating a valid workflow plan from a named toolchain is largely solved, whereas inferring an appropriate toolchain from biological intent alone is uniformly difficult regardless of model tier, compressing all models into a narrow 44-57% pass-rate band. Second, ablation shows that the dependency-structured plan and a completeness-reflection step drive performance, while adding a same-context validator-driven retry makes structural quality worse. Third, fault recovery and data-grounded interpretation remain unsolved. Models frequently propose fixes that force a clean exit while leaving the underlying data invalid, and data-grounded interpretation lags internal-knowledge recall by a consistent margin. Safety does not emerge from capability, and reasoning-tier models were among the least reliable at recognising unrecoverable faults. Once planning saturates, agent architecture and refusal calibration, not model scale, are the productive frontier.

19.
arXiv (CS.CV) 2026-06-19

LEAP: Layer-skipping Efficiency via Adaptive Progression for Vision Transformer Distillation

Vision Foundation Models (VFMs) with Vision Transformer (ViT) backbones, such as DINOv2, have become essential for downstream tasks like object recognition and semantic segmentation. The immense computational requirements of backbones often necessitate distillation into smaller architectures for edge deployment. Feature-based knowledge distillation (KD) often suffers from the teacher-student gap; the student struggles to imitate teacher's complex feature map due to its limited capacity. To mitigate this bottleneck, we propose LEAP: Layer-skipping Efficiency via Adaptive Progression, a training curriculum for ViT feature-based knowledge distillation. By utilizing the teacher's intermediate feature maps as a sequence of progressively more difficult targets, our curriculum allows the student to build a foundational representation before tackling higher-level abstractions. Our results demonstrate that this paradigm significantly accelerates convergence through adaptive difficulty selection across various student model sizes and dataset scales. With our curriculum, the LEAP-distilled ViT-S achieves 90.1% accuracy on ImageNet-100, a +12.24% improvement compared with baseline. On ImageNet-1K, LEAP achieves +3.84% and +7.75% improvement for the instance retrieval task on the Oxford and Paris datasets, respectively. Furthermore, the curriculum enables 25.1% savings in training FLOPs and 21% savings in training time on ImageNet-100 by implementing early-stopping for teacher inference during the initial stages of training. Code is available at https://github.com/KevinZ0217/LEAP

20.
arXiv (CS.CL) 2026-06-19

Where Does Social Reasoning Come From? Capability Provenance in Language Models

We use training-data attribution as an interpretable tool for capability discovery, mapping which regions of the pretraining corpus support social-reasoning versus STEM-reasoning in OLMo3-7B. Training-data attribution measures how strongly each training document influences a model's predictions on a benchmark, but document-level scores are too noisy to identify which corpus regions support which capabilities, and prior work has emphasized factual knowledge rather than reasoning. We compute gradient-based attribution (TrackStar via Bergson) over a working set drawn from the de-duplicated Dolma3 mix, aggregate influence across WebOrganizer's 24-format x 24-topic taxonomy (576 bins), and contrast benchmark pairs in a 2x2 design that varies domain (social vs. STEM) and capability type (reasoning vs. knowledge): SocialIQA and MMLU Social Sciences against ARC-Challenge and MMLU STEM. Social and STEM reasoning draw on qualitatively distinct corpus regions, and the contrast is sharper at the reasoning level than at the knowledge level. Targeted machine unlearning provides partial causal validation: forgetting high-attribution topic bins (e.g., Literature for SocialIQA) degrades the aligned benchmark more than within-bin random baselines, and we open-source all code, sampling manifests, the bin-level influence matrix, and unlearning checkpoints.

21.
medRxiv (Medicine) 2026-06-12

Integrative Mechanisms of Early Clinical and Research Training (ECART) in Orthopaedic Medical Education: A Qualitative Single-Case Study

Background: Early clinical exposure and student participation in research are important components of medical training. They may support learning motivation, evidence literacy, and self-directed learning. In many programmes, however, clinical training and research training remain separated. Few studies have explained, within a real teaching team, how learners turn clinical phenomena into researchable questions and how research participation can reshape their clinical understanding. Early Clinical and Research Training (ECART) is a clinical-research integration approach developed by an orthopaedic team at the Second Hospital of Shandong University. Methods: We conducted a theory-informed, interpretivist qualitative single-case study. The case was an orthopaedic clinical-research team at the Second Hospital of Shandong University. Participants included medical undergraduates, academic degree graduate students, professional degree graduate students, clinical teachers, and research platform leads. We used purposive sampling with maximum variation. Data were collected through semi-structured interviews and de-identified teaching documents. Data were analysed using the framework method and were interpreted with a Context-Activity-Mechanism-Outcome (CAMO) logic. Results: The analysis showed that ECART was not simply early entry into the clinic or early entry into the laboratory. It was a team-based learning process centred on real medical problems. Four themes were identified. First, early clinical exposure helped learners make real problems visible and nameable, rather than merely increasing exposure. Second, clinical-research connection followed different pathways. Professional degree graduate students often started from clinical uncertainties in residency training and case management, and moved toward evidence-informed small projects. Academic degree graduate students often started from literature gaps, experimental findings, and mechanistic hypotheses, and then used clinical feedback to calibrate meaning. Third, research training, through literature reading, group meetings, experimental design, data review, and mentor questioning, helped learners move from completing tasks to explaining problems. Fourth, sustained ECART depended on a tiered team ecology formed by clinical teachers, research mentors, research platforms, and senior peers. Based on these findings, we refined the ECART programme theory: real medical problems are translated through explanation, searching, experimentalisation, and feedback-based reinterpretation into research questions that learners can understand, discuss, and test. This process supports problem formation, evidence awareness, mechanistic reasoning, translational judgement, and career clarification. Conclusion: ECART is best understood as a clinical-research integrated learning ecology that emerges from real team practice, rather than as a fixed standardised course. Its educational value lies in a recurring cycle of real problems, research translation, multi-source feedback, and clinical reinterpretation. This framework may inform the design, evaluation, and contextual adaptation of clinical-research integration pathways in medical education.

22.
arXiv (CS.CL) 2026-06-15

SuperThoughts: Reasoning Tokens in Superposition

Long Chain-of-Thought (CoT) reasoning improves LLM problem-solving but is computationally expensive due to sequential token generation. While recent works explore reasoning in continuous latent spaces to bypass discrete token generation, they often struggle with training stability and fail to scale to complex, long-horizon tasks due to lack of supervision signal. We propose SuperThoughts, which compresses pairs of consecutive CoT tokens into single latent representations and decodes two tokens per step via a lightweight Multi-Token Prediction (MTP) module. This preserves discrete token supervision at training time while doubling throughput at inference time. We finetune Qwen2.5-Math-1.5B-Instruct, Qwen2.5-Math-7B-Instruct, Qwen2.5-Math-14B-Instruct, and evaluate on MATH500, AMC, OlympiadBench, and GPQA-Diamond. With a confidence-based adaptive mechanism that falls back to standard decoding when uncertain, SuperThoughts achieves $\sim$20–30\% CoT length reduction while maintaining accuracy with minimal degradation (1-2 points accuracy drop on most tasks).

23.
arXiv (CS.CV) 2026-06-12

Iterative Tool Usage Exploration for Multimodal Agents via Step-wise Preference Tuning

Multimodal agents, which integrate a controller e.g., a vision language model) with external tools, have demonstrated remarkable capabilities in tackling complex multimodal tasks. Existing approaches for training these agents, both supervised fine-tuning and reinforcement learning, depend on extensive human-annotated task-answer pairs and tool trajectories. However, for complex multimodal tasks, such annotations are prohibitively expensive or impractical to obtain. In this paper, we propose an iterative tool usage exploration method for multimodal agents without any pre-collected data, namely SPORT, via step-wise preference optimization to refine the trajectories of tool usage. Our method enables multimodal agents to autonomously discover effective tool usage strategies through self-exploration and optimization, eliminating the bottleneck of human annotation. SPORT has four iterative components: task synthesis, step sampling, step verification, and preference tuning. We first synthesize multimodal tasks using language models. Then, we introduce a novel trajectory exploration scheme, where step sampling and step verification are executed alternately to solve synthesized tasks. In step sampling, the agent tries different tools and obtains corresponding results. In step verification, we employ a verifier to provide AI feedback to construct step-wise preference data. The data is subsequently used to update the controller for tool usage through preference tuning, producing a SPORT agent. By interacting with real environments, the SPORT agent gradually evolves into a more refined and capable system. Evaluation in the GTA and GAIA benchmarks shows that the SPORT agent achieves 6.41% and 3.64% improvements, underscoring the generalization and effectiveness introduced by our method. The project page is https://SPORT-Agents.github.io.

24.
arXiv (CS.LG) 2026-06-16

Near-Optimal Regret for Distributed Adversarial Bandits: A Black-Box Approach

arXiv:2602.06404v2 Announce Type: replace Abstract: We study distributed adversarial bandits, where $N$ agents cooperate to minimize the global average loss while observing only their own local losses. We show that the minimax regret for this problem is $\tilde{\Theta}(\sqrt{(\rho^{-1/2}+K/N)T})$, where $T$ is the horizon, $K$ is the number of actions, and $\rho$ is the spectral gap of the communication matrix. Our algorithm, based on a novel black-box reduction to bandits with delayed feedback, requires agents to communicate only through gossip. It achieves an upper bound that significantly improves over the previous best bound $\tilde{O}(\rho^{-1/3}(KT)^{2/3})$ of Yi and Vojnovic (2023). We complement this result with a matching lower bound, showing that the problem's difficulty decomposes into a communication cost $\rho^{-1/4}\sqrt{T}$ and a bandit cost $\sqrt{KT/N}$. We further demonstrate the versatility of our approach by deriving first-order and best-of-both-worlds bounds in the distributed adversarial setting. Finally, we extend our framework to distributed linear bandits in $R^d$, obtaining a regret bound of $\tilde{O}(\sqrt{(\rho^{-1/2}+1/N)dT})$, achieved with only $O(d)$ communication cost per agent and per round via a volumetric spanner.

25.
arXiv (CS.AI) 2026-06-18

DRIFT: Refining Instruction Data via On-Policy Data Attribution

arXiv:2606.18307v1 Announce Type: cross Abstract: Optimizing the training data distribution for Supervised Fine-Tuning (SFT) dictates the capability of Large Language Models (LLMs). While existing data curation methods excel at accelerating training under constrained budgets, they are less suited to elevating the capability upper bound. The challenge here is no longer to identify a smaller subset that preserves performance, but to refine the data distribution toward instances most capable of improving the final model. To address this problem, we explore instance-level data attribution using Influence Functions (IF). We identify that standard IF formulations struggle in this setting due to two structural limitations: a proximity gap caused by off-policy validation targets, and a severe bias towards gradient norm. We propose DRIFT (Data Refinement via On-Policy Influence Functions for Supervised Fine-Tuning). Instead of relying on external reference data, DRIFT utilizes the model's on-policy rollouts as validation targets, which empirically minimizes the parameter proximity gap and better aligns with the local neighborhood assumption of IF. It further applies signed weighting based on trajectory correctness and debiases influence scores against the gradient hacking issue, allowing a small set of validation queries to act as reliable anchors for attributing the full dataset. Experiments on 7B-parameter instruction and reasoning models show that DRIFT consistently raises the performance ceiling on both, outperforming existing data curation baselines.