On-Policy Distillation (OPD) has emerged as a widely adopted post-training paradigm for enhancing large language models in reasoning domains. However, the data-centric mechanisms in OPD remain relatively underexplored. This paper presents a empirical study of data efficiency and data selection in OPD. We begin by investigating an extreme setting: training OPD on only one example, namely 1-shot OPD. Surprisingly, we find that 1-shot OPD is consistently effective across all sampled training examples and harder examples often yield superior performance gain. We next investigate what actually drives the student model's improvement in the training data. Our analysis reveals that the improvement is not driven by high token entropy, but the longer CoT paths which hard problems naturally generate. Training on longer CoT can help maintain closer alignment with the teacher over a long reasoning horizon, and learn critical thinking patterns usually missing in short CoTs, such as reflection (e.g., ``Alternatively''). Based on these insights, we propose a simple data selection method that selects only hard examples for training, where even ``unsolvable'' examples that completely exceed the teacher's capability can be successfully used. Our experiments conducted on four models ranging from 1.5B to 7B show that training the student model on only 8 selected hard examples matches the performance of the 17K dataset baseline.
Zhinan Hou, Jiaqi Zhang, Xunliang Cai et al.· 0 citations
In this paper we illustrate a novel architecture generating interpretable behavior and explanations. We refer to this architecture as the Mirror Agent Model because it defines the observer model, that is the target of explicit and implicit communications, as a mirror of the agent's. With the goal of providing a general understanding of this work, we firstly show prior relevant results addressing the informative communication of agents intentions and the production of legible behavior. In the second part of the paper we furnish the architecture with novel capabilities for explanations through off-the-shelf saliency methods, followed by preliminary qualitative results.
This study introduces an intelligent framework that integrates machine learning and deep neural network ensemble techniques for early detection and prognosis of cardiovascular diseases. The system utilizes real-time physiological data collected from Internet of Medical Things (IoMT) devices, including ECG sensors, heart rate monitors, and blood pressure trackers. To ensure the accuracy and reliability of input data, preprocessing steps such as noise reduction, normalization, and missing value imputation are employed. The most significant health indicators are identified through effective feature selection methods and then processed using optimized classifiers such as Support Vector Machines (SVM), Random Forests, and eXtreme Gradient Boosting (XGBoost), which are combined in an ensemble architecture to improve diagnostic precision. The framework demonstrates remarkable performance in predicting cardiovascular disease risk, achieving higher accuracy, reduced false positives, and enhanced consistency compared to conventional methods. It is designed on a cloud-based infrastructure that ensures scalability and real-time processing for continuous patient monitoring. Experimental evaluation on real-world cardiovascular datasets confirms the framework's efficiency in early-stage risk assessment and clinical decision support. The results highlight the potential of combining traditional machine learning and deep learning paradigms to achieve proactive healthcare management and improve patient outcomes.
Scientific papers require models to reason jointly over text, equations, figures, tables, code, and datasets while preserving the provenance of supporting evidence. Existing benchmarks typically evaluate these capabilities in isolation, leaving unclear whether multimodal models can support realistic scientific-reading workflows. We introduce SciDocBench, a workflow-centered benchmark for scientific document understanding. It contains 124 expert-authored and difficulty-screened questions organized into seven research-assistant capability groups and 19 subtasks across five scientific domains. Each question is instantiated under four matched conditions combining English or Chinese questions with all-images-first or interleaved document representations, yielding 496 evaluation instances for controlled analysis. The strongest evaluated system achieves only 62.6/100, with pronounced weaknesses in document perception, evidence grounding, verification, and cross-document reasoning. To translate these diagnostics into scalable training signals, we introduce SciDocIR, a typed evidence-graph representation that preserves scientific document objects, layout and cross-reference relations, and provenance. Building on SciDocIR, we construct SciDocDataset, comprising approximately 15K supervised fine-tuning samples and 8K reinforcement-learning samples across 14 verifiable subtasks. Together, SciDocBench, SciDocIR, and SciDocDataset form an evaluation-to-training framework for diagnosing and improving scientific-document assistants. The project page is available at https://github.com/InternLM/SciDocBench.
Shenxi Wu, Yuhong Liu, Haosong Zhang et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Large language models are now trained and evaluated under a diverse set of paradigms: supervised fine-tuning (SFT), few-shot in-context learning (ICL), KL-regularized RLHF/RLVR, on-policy distillation (OPD), and test-time reasoning with search and chain-of-thought. These methods are often discussed as fundamentally different, and recent empirical results--such as the mixed impact of few-shot prompting on RL-tuned reasoning models--can appear puzzling. This note develops a Bayesian perspective that puts these procedures on the same footing. At the core is a two-step template: (i) construct a (generalized) Bayes or Gibbs posterior q* over outputs or actions given a context, using a prior/reference model and a utility signal (log-likelihood, reward, or advantage); and (ii) approximate q* by a forward-KL projection onto a parametric family, either in-weights (SFT/RL) or in-context (ICL). Part I formalizes few-shot ICL and SFT as amortized and-weights projections onto the Bayes posterior predictive. Parts II-IV show that KL-regularized RLHF/RLVR, reward-weighted SFT, reward-weighted ICL (RW-ICL), and advantage-weighted SFT (AWSFT) are all instances of forward-KL projection onto posteriors induced by rewards or advantages. We disentangle where these equivalences hold (objectives and first-order updates) and where they do not (source and granularity of the learning signal). Part V sketches implications for modern reasoning pipelines: RLHF/RLVR recipes as "posterior design + projection", why cold-start or supervised warm-up is practically unavoidable for importance-weighted KL projections, and DeepSeek-R1 and o1-style reasoning models as combining test-time Bayesian search with training-time KL amortization.
Biological agents navigate familiar environments not by re-solving routes for each new goal, but by reusing a learned map built once and read off as goals change. Existing artificial cognitive-map models mimic this reuse, yet their guidance is not explicitly grounded in additive heterogeneous route costs. Furthermore, they often struggle with memory efficiency: representative state-indexed and high-rank spectral constructions incur substantial storage growth as the environment scales. We present BCM, which grounds a reusable cognitive map in local edge costs through a self-supervised Bellman-grounded objective and a compact coordinate encoding, supporting changing goal queries without per-goal retraining. On weighted grids of up to $N=1600$ nodes, BCM maintains full success and only a 5\% mean Gap relative to exact Dijkstra search, compared with about $45\%$ for a connectivity-based spectral baseline. Notably, as the graph size increases from $N=400$ to $N=3600$, its memory footprint grows sublinearly while maintaining competitive performance, making our method scalable to complex environments. Together, these results show that additive route costs can be written into a compact, reusable cognitive-map representation, bridging the gap between biological flexibility and optimal path planning.
Electroencephalogram (EEG) visual decoding aims to recover visual semantics from non-invasive neural time-series signals, for which robust alignment between noisy neural responses and stable semantic representations is key to achieving high-performance decoding. Despite recent advances in contrastive learning, robust EEG decoding remains challenging because existing methods rely on fixed visual or textual anchors whose semantic relations may become misaligned with EEG representations that vary across trials, subjects, and learning stages. Our empirical evidence shows that this instability appears across both standard EEG decoding protocols and more challenging robustness settings, including strict cross-subject transfer and realistic personalized continual adaptation. We provide a formal analysis showing that fixed semantic supervision can bias optimization when EEG-specific relations evolve, and that structure-agnostic perturbations may distort semantically important EEG components. To address these issues, we propose Progressive Contrastive Alignment (ProCA), a unified and model-agnostic framework for adaptive neural-semantic alignment. ProCA progressively refines class-level contrastive supervision from frozen vision-language priors to EEG-aware semantic relations, and introduces structure-consistent interpolation to constrain feature mixing according to channel-wise and temporal importance. Across subject-dependent, subject-independent, strict cross-subject transfer, and continual adaptation settings, ProCA achieves average relative Top-1/Top-5 gains of 7.4%/3.9%, 10.0%/4.6%, 28.1%/17.8%, and 16.8%/11.6%, respectively.
Kanglei Zhou, Chunyan Lan, Dongyang Li et al.· 0 citations
We present Discovery Loop, a lightweight system that uses a large language model (LLM) to iteratively evolve optimization algorithms. Starting from a simple seed solver, the LLM proposes algorithmic improvements guided by a scoreboard of results and a history of prior ideas. Each candidate is evaluated against an independent verifier; improvements are kept and failures discarded. Applied to the Packomania circle-packing benchmark (csqv: maximize the sum of radii of N variable-radius circles in the unit square), the system improved the best known solutions for 10 values of N in the range 101-114, with gains of 2.4%-5.4% over prior records, all within 15 iterations and at a total LLM cost of $27.72. These results have been independently accepted by Packomania. We describe the method, analyze cost-efficiency dynamics including an adaptive plateau-detection mechanism, and discuss implications for democratizing automated scientific discovery.
Evaluation of medical artificial intelligence agents remains predominantly answer-centric, assessing only the correctness of final outputs while overlooking the quality of intermediate reasoning. In clinical settings, however, a correct answer reached through fabricated evidence or incoherent logic is as dangerous as an incorrect one. We propose MedTraj, a framework that treats reasoning trajectories as critical objects for construction, evaluation, and optimization. The pipeline generates structured multi-step reasoning chains from medical reasoning sources. Each trajectory is then parsed into clinical observations, evidence, numbered reasoning steps, and a final conclusion, and scored across five quality dimensions: coherence, evidence support, hallucination, completeness, and traceability. Controlled error injection introduces targeted faults into otherwise correct trajectories to establish causal links between specific reasoning failures and measurable quality degradation. Building on this, step-level filtering based on marginal contribution identifies which individual reasoning steps drive or undermine trajectory quality. Finally, quality-weighted context learning feeds trajectory evaluations back into the model at inference time, allowing it to learn from both strong and weak reasoning demonstrations. Experiments across CareQA, PubMedQA, and CECMed demonstrate that trajectory context consistently improves reasoning coherence, with gains of +0.029 to +0.041 over a zero-shot baseline. On CECMed, quality-weighted context nearly doubles the correctness over the zero-shot baseline while cutting the hallucination ratio by 87%. Marginal-contribution analysis further shows that a small minority of reasoning steps carry most of the quality signal, and that extending chains beyond four steps yields diminishing returns.
AI oversight methods rely on ground truth for validation, but what constitutes appropriate AI behavior is contested. This leaves evaluation of moral reasoning in LLMs and debate-based oversight implicitly avoiding realistic ambiguity. We investigate an alternative standard designed to function despite such ambiguity: structural quality of the defence a model can mount for its verdicts in response to critical questions, measured through a four-phase dialectical protocol grounded in Walton's theory of argumentation schemes and Govier's criteria for argument cogency. The protocol is adaptive to different frames of reasoning, extends beyond multiple-choice framing, and treats both the reasoning that precedes a verdict and its post-hoc justification. Across nine frontier models and 200 high-ambiguity MoralChoice items -- $6,778$ judge-scored cells, validated against $89.6\%$ inter-judge agreement on the binary failure judgment -- models defend their reasoning well above the rubric minimum on every dimension. Failure mass concentrates on grounds and sufficiency, and correlates with epistemic hedging rather than argument length. Reasoning is better defended than post-hoc justification, on every model and every Govier dimension. The scheme a model presents in its justification differs from the one it reasoned with on a substantial share of dilemmas ($\geq 20\%$ per model), despite value-based practical reasoning dominating both tracks. The protocol catches strictly indefensible defences (self-contradiction, false premises), and it surfaces difficulties in characterizing the role of retraction in AI alignment, suggesting a need for more situated evaluations.
Daan R. Henselmans, Derck W. E. Prinzhorn, Arno Libert· 1 citation
Autonomous coding agents are increasingly proposed as AI-scientist systems that conduct analyses and write research reports, but executing a prescribed analysis is not the same as making a discovery. Existing benchmarks are configured for reproduction: tasks, data, and rubrics are built around a hidden target study, and recovery of its result is rewarded. We present TruthInsightBench, a benchmark configured for discovery. Its 40 blind tasks, drawn from 40 peer-reviewed studies across 10 scientific domains, expose only a neutral scientific objective and frozen data; source conclusions, expected values, and analysis paths are withheld, leaving the agent to determine what claim the data support. A fixed LLM-based judge scores the evidentiary maturity of an agent's own claims along six dimensions, operationalized as 29 artifact-grounded items, with automated, deterministic aggregation and no per-instance human grading, so evaluation can be repeated automatically as agents evolve. On one frozen base model, four coding agents form a narrow plateau (58.4-60.3 of 100) with no statistically reliable pairwise separation: they execute and document analyses competently, with comparatively strong evidence auditability and novelty, but largely lack the discriminating acts that establish a trustworthy claim (controls, robustness, falsifiability, and cross-dataset generalization). The bottleneck is scientific judgment rather than coding, and genuine discovery remains out of reach. TruthInsightBench makes this gap a measurable target; data and scoring code are at https://github.com/TruthInsight-stack/TruthInsightBench.
Zhibo Yang, Chen Zhang, Yuewei Zhang et al.· 0 citations
General continual learning (GCL) aims to learn from evolving data streams without task identities, explicit boundaries, or repeated access to previous data, making it a realistic yet challenging setting for continual intelligence. Although pretrained models (PTMs) provide rich prior knowledge for addressing the limited supervision and non-stationary nature of GCL, existing PTM-based methods often directly adapt pretrained representations and overlook two critical gaps: the misalignment between upstream pretraining and downstream continual adaptation, and the unreliability of conventional output alignment under blurry streams. Here we propose MePo++, a unified post-training framework that bridges pretrained knowledge and downstream GCL through representation refinement and reconciliation. MePo++ introduces two complementary components: MetaPrep, which improves representation plasticity for continual adaptation through unsupervised meta-refinement over pseudo continual sequences; and StreamAlign, which reinforces representation stability by reconciling evolving online features with a stable pretrained geometry. By improving representation learnability before adaptation and preserving alignment during continual learning, MePo++ enables PTMs to remain both plastic for new concepts and stable over evolving streams. Experiments across diverse PTMs, datasets, and continual learning baselines demonstrate the consistent effectiveness and generality of MePo++ for PTM-based GCL. Our code is available at https://github.com/SunGL001/MePo_Plus.
Guanglong Sun, Kanglei Zhou, Liyuan Wang et al.· 0 citations
