Committee

Gemini Diffusion is also useful for tasks such as refactoring code, adding new features to applications, or converting an existing codebase to a different language. Read More

Ultimately, the big takeaway for ML researchers is that before proclaiming an AI milestone—or obituary—make sure the test itself isn’t flawedRead More

Patients using chatbots to assess their own medical conditions may end up with worse outcomes than conventional methods, according to a new Oxford study.Read More

The Growing Complexity of Reasoning Data Curation Recent reasoning models, such as DeepSeek-R1 and o3, have shown outstanding performance in mathematical, coding, and scientific areas, utilizing post-training techniques like supervised fine-tuning (SFT) and reinforcement learning (RL). However, the complete methodologies behind these frontier reasoning models are not public, which makes research for building reasoning models difficult. While SFT data curation has become a powerful approach for developing strong reasoning capabilities, most existing efforts explore only limited design choices, such as relying solely on human-written questions or single teacher models. Moreover, exploring the extensive design space of various techniques for generating question-answer pairs requires high costs for teacher inference and model training. Reasoning traces provided by models such as Gemini, QwQ, and DeepSeek-R1 have enabled knowledge distillation techniques to train smaller reasoning models. Projects like OpenR1, OpenMathReasoning, and OpenCodeReasoning collect questions from public forums and competition sites, while Natural Reasoning utilizes pre-training corpora as seed data. Some efforts, such as S1 and LIMO, focus on manually curating small, high-quality datasets of challenging prompts. Other methods, such as DeepMath-103K and Nvidia Nemotron, introduce innovations across data sourcing, filtering, and scaling stages. RL methods, including AceReason and Skywork-OR1, have enhanced reasoning capabilities beyond traditional SFT methods. OpenThoughts: A Scalable Framework for SFT Dataset Development Researchers from Stanford University, the University of Washington, BespokeLabs.ai, Toyota Research Institute, UC Berkeley, and 12 additional organizations have proposed OpenThoughts, a new SOTA open reasoning data recipe. OpenThoughts uses a progressive approach across three iterations: OpenThoughts-114K scales the Sky-T1 pipeline with automated verification, OpenThoughts2-1M enhances data scale through augmented question diversity and synthetic generation strategies, and OpenThoughts3-1.2M incorporates findings from over 1,000 ablation experiments to develop a simple, scalable, and high-performing data curation pipeline. Moreover, the model OpenThinker3-7B achieves state-of-the-art performance among open-data models at the 7B scale. The OpenThoughts3-1.2M is built by ablating each pipeline component independently while maintaining constant conditions across other stages, generating 31,600 data points per strategy and fine-tuning Qwen2.5-7B-Instruct on each resulting dataset. The goal during training is to create the best dataset of question-response pairs for SFT reasoning. Evaluation occurs across eight reasoning benchmarks across mathematics (AIME24, AMC23, MATH500), coding (CodeElo, CodeForces, LiveCodeBench), and science (GPQA Diamond, JEEBench). The experimental design includes a rigorous decontamination process to remove high-similarity samples and maintains a held-out benchmark set for generalization testing. Evalchemy serves as the primary evaluation tool, ensuring consistent evaluation protocols. Evaluation Insights and Benchmark Performance The OpenThoughts pipeline evaluation reveals key insights across question sourcing, mixing, filtering, answer filtering, and the teacher model. Question sourcing experiments show that CodeGolf and competitive coding questions achieve the highest performance for code tasks (25.3-27.5 average scores), while LLM-generated and human-written questions excel in mathematics (58.8-58.5 scores), and physics StackExchange questions with chemistry textbook extractions perform best in science (43.2-45.3 scores). Mixing question shows that combining multiple question sources degrades performance, with optimal results of 5% accuracy improvements over diverse mixing strategies. In the teacher model, QwQ-32B outperforms DeepSeek-R1 in knowledge distillation, achieving an accuracy improvement of 1.9-2.6%. In conclusion, researchers present the OpenThoughts project, showing that systematic experimentation can significantly advance SFT data curation for reasoning models. Researchers developed OpenThoughts3-1.2M, a state-of-the-art open-data reasoning dataset across science, mathematics, and coding domains. The resulting OpenThinker3-7B model achieves superior performance among open-data reasoning models at its scale. However, several limitations remain unexplored, including RL approaches, staged fine-tuning, and curriculum learning strategies. Future research directions include investigating cross-domain transfer effects when optimizing individual domains versus overall performance, and understanding the scaling dynamics as student models approach teacher capabilities. Check out the Paper, Project Page and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 99k+ ML SubReddit and Subscribe to our Newsletter. The post OpenThoughts: A Scalable Supervised Fine-Tuning SFT Data Curation Pipeline for Reasoning Models appeared first on MarkTechPost.

arXiv:2506.09277v2 Announce Type: replace Abstract: Large Language Models (LLM) have demonstrated the capability of generating free text self Natural Language Explanation (self-NLE) to justify their answers. Despite their logical appearance, self-NLE do not necessarily reflect the LLM actual decision-making process, making such explanations unfaithful. While existing methods for measuring self-NLE faithfulness mostly rely on behavioral tests or computational block identification, none of them examines the neural activity underlying the model’s reasoning. This work introduces a novel flexible framework for quantitatively measuring the faithfulness of LLM-generated self-NLE by directly comparing the latter with interpretations of the model’s internal hidden states. The proposed framework is versatile and provides deep insights into self-NLE faithfulness by establishing a direct connection between self-NLE and model reasoning. This approach advances the understanding of self-NLE faithfulness and provides building blocks for generating more faithful self-NLE.

Artificial intelligence has undergone a significant transition from basic language models to advanced models that focus on reasoning tasks. These newer systems, known as Large Reasoning Models (LRMs), represent a class of tools designed to simulate human-like thinking by producing intermediate reasoning steps before arriving at conclusions. The focus has moved from generating accurate outputs to understanding the process that leads to these answers. This shift has raised questions about how these models manage tasks with layered complexity and whether they truly possess reasoning abilities or are simply leveraging training patterns to guess outcomes. Redefining Evaluation: Moving Beyond Final Answer Accuracy A recurring problem with evaluating machine reasoning is that traditional benchmarks mostly assess the final answer without examining the steps involved in arriving at it. Final answer accuracy alone does not reveal the quality of internal reasoning, and many benchmarks are contaminated with data that may have been seen during training. This creates a misleading picture of a model’s true capabilities. To explore actual reasoning, researchers require environments where problem difficulty can be precisely controlled and intermediate steps can be analyzed. Without such settings, it is hard to determine whether these models can generalize solutions or merely memorize patterns. To evaluate reasoning more reliably, the research team at Apple designed a setup using four puzzle environments: Tower of Hanoi, River Crossing, Checkers Jumping, and Blocks World. These puzzles allow precise manipulation of complexity by changing elements such as the number of disks, checkers, or agents involved. Each task requires different reasoning abilities, such as constraint satisfaction and sequential planning. Importantly, these environments are free from typical data contamination, enabling thorough checks of both outcomes and the reasoning steps in between. This method ensures a detailed investigation of how models behave across varied task demands. The research introduced a comparative study using two sets of models: Claude 3.7 Sonnet and DeepSeek-R1, along with their “thinking” variants and their standard LLM counterparts. These models were tested across the puzzles under identical token budgets to measure both accuracy and reasoning efficiency. This helped reveal performance shifts across low, medium, and high-complexity tasks. One of the most revealing observations was the formation of three performance zones. In simple tasks, non-thinking models outperformed reasoning variants. For medium complexity, reasoning models gained an edge, while both types collapsed completely as complexity peaked. Comparative Insights: Thinking vs. Non-Thinking Models Under Stress An in-depth analysis revealed that reasoning effort increased with task difficulty up to a certain point but then declined despite the availability of resources. For instance, in the Tower of Hanoi, Claude 3.7 Sonnet (thinking) maintained high accuracy until complexity reached a certain threshold, after which performance dropped to zero. Even when these models were supplied with explicit solution algorithms, they failed to execute steps beyond specific complexity levels. In one case, Claude 3.7 could manage around 100 steps correctly for the Tower of Hanoi but was unable to complete simpler River Crossing tasks requiring only 11 moves when $N = 3$. This inconsistency exposed serious limitations in symbolic manipulation and exact computation. The performance breakdown also highlighted how LRMs handle their internal thought process. Models frequently engaged in “overthinking,” generating correct intermediate solutions early in the process but continuing to explore incorrect paths. This led to inefficient use of tokens. At medium complexity levels, models began to find correct answers later in their reasoning chains. However, at high levels of complexity, they failed to produce accurate solutions. Quantitative analysis confirmed that solution accuracy dropped to near zero as the problem complexity increased, and the number of reasoning tokens allocated began to decline unexpectedly. Scaling Limits and the Collapse of Reasoning This research presents a sobering assessment of how current Learning Resource Management Systems (LRMs) operate. Research from Apple makes it clear that, despite some progress, today’s reasoning models are still far from achieving generalized reasoning. The work identifies how performance scales, where it collapses, and why over-reliance on benchmark accuracy fails to capture deeper reasoning behavior. Controlled puzzle environments have proven to be a powerful tool for uncovering hidden weaknesses in these systems and emphasizing the need for more robust designs in the future. Check out the Paper. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 99k+ ML SubReddit and Subscribe to our Newsletter. The post Apple Researchers Reveal Structural Failures in Large Reasoning Models Using Puzzle-Based Evaluation appeared first on MarkTechPost.

arXiv:2506.10617v1 Announce Type: cross Abstract: This paper addresses the persistent challenge of accurately digitizing paper-based electrocardiogram (ECG) recordings, with a particular focus on robustly handling single leads compromised by signal overlaps-a common yet under-addressed issue in existing methodologies. We propose a two-stage pipeline designed to overcome this limitation. The first stage employs a U-Net based segmentation network, trained on a dataset enriched with overlapping signals and fortified with custom data augmentations, to accurately isolate the primary ECG trace. The subsequent stage converts this refined binary mask into a time-series signal using established digitization techniques, enhanced by an adaptive grid detection module for improved versatility across different ECG formats and scales. Our experimental results demonstrate the efficacy of our approach. The U-Net architecture achieves an IoU of 0.87 for the fine-grained segmentation task. Crucially, our proposed digitization method yields superior performance compared to a well-established baseline technique across both non-overlapping and challenging overlapping ECG samples. For non-overlapping signals, our method achieved a Mean Squared Error (MSE) of 0.0010 and a Pearson Correlation Coefficient (rho) of 0.9644, compared to 0.0015 and 0.9366, respectively, for the baseline. On samples with signal overlap, our method achieved an MSE of 0.0029 and a rho of 0.9641, significantly improving upon the baseline’s 0.0178 and 0.8676. This work demonstrates an effective strategy to significantly enhance digitization accuracy, especially in the presence of signal overlaps, thereby laying a strong foundation for the reliable conversion of analog ECG records into analyzable digital data for contemporary research and clinical applications. The implementation is publicly available at this GitHub repository: https://github.com/masoudrahimi39/ECG-code.

arXiv:2506.10737v1 Announce Type: new Abstract: The rapid evolution of scientific fields introduces challenges in organizing and retrieving scientific literature. While expert-curated taxonomies have traditionally addressed this need, the process is time-consuming and expensive. Furthermore, recent automatic taxonomy construction methods either (1) over-rely on a specific corpus, sacrificing generalizability, or (2) depend heavily on the general knowledge of large language models (LLMs) contained within their pre-training datasets, often overlooking the dynamic nature of evolving scientific domains. Additionally, these approaches fail to account for the multi-faceted nature of scientific literature, where a single research paper may contribute to multiple dimensions (e.g., methodology, new tasks, evaluation metrics, benchmarks). To address these gaps, we propose TaxoAdapt, a framework that dynamically adapts an LLM-generated taxonomy to a given corpus across multiple dimensions. TaxoAdapt performs iterative hierarchical classification, expanding both the taxonomy width and depth based on corpus’ topical distribution. We demonstrate its state-of-the-art performance across a diverse set of computer science conferences over the years to showcase its ability to structure and capture the evolution of scientific fields. As a multidimensional method, TaxoAdapt generates taxonomies that are 26.51% more granularity-preserving and 50.41% more coherent than the most competitive baselines judged by LLMs.

arXiv:2506.10314v1 Announce Type: cross Abstract: Malicious sockpuppet detection on Wikipedia is critical to preserving access to reliable information on the internet and preventing the spread of disinformation. Prior machine learning approaches rely on stylistic and meta-data features, but do not prioritise adaptability to author-specific behaviours. As a result, they struggle to effectively model the behaviour of specific sockpuppet-groups, especially when text data is limited. To address this, we propose the application of meta-learning, a machine learning technique designed to improve performance in data-scarce settings by training models across multiple tasks. Meta-learning optimises a model for rapid adaptation to the writing style of a new sockpuppet-group. Our results show that meta-learning significantly enhances the precision of predictions compared to pre-trained models, marking an advancement in combating sockpuppetry on open editing platforms. We release a new dataset of sockpuppet investigations to foster future research in both sockpuppetry and meta-learning fields.

arXiv:2506.10406v1 Announce Type: new Abstract: Large Language Models (LLMs) have demonstrated impressive capabilities in complex reasoning tasks, yet they still struggle to reliably verify the correctness of their own outputs. Existing solutions to this verification challenge often depend on separate verifier models or require multi-stage self-correction training pipelines, which limit scalability. In this paper, we propose Policy as Generative Verifier (PAG), a simple and effective framework that empowers LLMs to self-correct by alternating between policy and verifier roles within a unified multi-turn reinforcement learning (RL) paradigm. Distinct from prior approaches that always generate a second attempt regardless of model confidence, PAG introduces a selective revision mechanism: the model revises its answer only when its own generative verification step detects an error. This verify-then-revise workflow not only alleviates model collapse but also jointly enhances both reasoning and verification abilities. Extensive experiments across diverse reasoning benchmarks highlight PAG’s dual advancements: as a policy, it enhances direct generation and self-correction accuracy; as a verifier, its self-verification outperforms self-consistency.