Training and data
39 notes, newest first. Includes the vault tags training, fine-tuning, synthetic data, datasets, dataset, training data, data.
A Privileged Teacher That Only Scores Cannot Search: Critique-Conditioned On-Policy Self-Distillation for Reasoning
The study tests whether scoring signals in on-policy self-distillation localize errors and push probability away from erroneous student tokens. It reports that full-distribution KL and token log-ratio signals do…
Use and Disuse: Intent-Structured Experience Consolidation for Memory and Learning in LLM Agents
The preprint proposes Hippocam, a hierarchical memory and continual-learning architecture that organizes ongoing work as nested intents. Active context stays centered on the current intent, while completed intents are…
GeoReform: Reflective Formalization Evolution for Multimodal Geometry Problem Solving
The preprint proposes GeoReform, which treats geometric formalization as an optimizable policy rather than a fixed parser output. It collects failed rollouts, diagnoses representation defects, and mutates how entities…
Recursive Self-Improvement through Multi-Agent Self-Supervision
The work describes MASS, which alternates evolutionary workflow optimization with supervised fine-tuning on self-generated trajectories. The abstract reports discovery of multi-agent workflows and, over two cycles with…
ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences
The preprint formalizes fixed-parameter program self-evolution that links task solving, scientific verification, and program updates. ScienceClaw-Eval spans 23 disciplines and measures correctness, evolutionary gain…
Sherpa: Teaching LLMs to Teach Adaptively
The preprint introduces Sherpa, a multi-turn reinforcement-learning framework that trains an LLM teacher against multiple student archetypes with distinct learning preferences. The teacher is optimized directly against…
VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning
The preprint proposes an agent harness that co-evolves the policy, training curriculum, and judge. A Policy Improvement Loop diagnoses recurring failures and builds an adaptive curriculum; when verification becomes a…
ImproveAnyTask: An Autonomous Post-Training Harness for Iterative Model Self-Improvement
The preprint describes an autonomous post-training harness that attributes errors, selects update directions, and executes model updates under a limited compute budget. It combines metric- and case-level analysis…
Credit Where It Matters: Dependency-Aware Policy Optimization for Terminal Agents
The preprint proposes Dependency-Aware Group Policy Optimization (DepGPO) for multi-step terminal tasks. It builds a command dependency graph from execution traces, traces backward from resources inspected by a…
Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis
The preprint introduces task-harness co-evolution for recursive reasoning-data synthesis. Online self-improvement turns solver failures into reusable skills, while post-task updates revise skills, prompts, and…
Agent Error Dataset: Scaling 50,000 Error--Diagnosis Pairs for Failure Analysis and Error-Aware Post-Training
The preprint introduces 50,228 error-diagnosis pairs from 9,961 tasks across 33 environments, 19 harness families, and 23 policy models. Its Agentic Error-to-Training pipeline preserves traces and execution metadata…
PivotOPD: Learning to Recover from Pivotal Mistakes in Multi-Turn Agents
The technical report studies pivotal mistakes that move an agent farther from task completion during multi-turn interaction. Across three Qwen3 models, the abstract reports that more than half of failed rollouts…
SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance
SAGE proposes Structural Admissibility-Guided Exploration for long-horizon reasoning under sparse rewards. It frames two failure modes—exploration bias toward locally plausible but unstable branches and compounding…
SciWalker: Synthesizing Scientific Coding Problems with Operator Graphs and Execution Feedback
The preprint describes a framework for synthesizing scientific coding problems by sampling operator chains from graphs built from scientific library interfaces and operation modes. An LLM generates problem statements…
Synthetic Hospital: An Open, Verifiable, Physician-Validated Longitudinal EHR Benchmark
The preprint describes a fully synthetic, fact-grounded longitudinal EHR benchmark built from public medical education material, with no PHI. The abstract reports 1,268 patients and 5,602 encounters, ICD-10-CM/SNOMED…
Self-Play Pretraining with Zero Data
The paper presents a proof-of-concept in which a generator and learner train in tandem without natural training data. The generator proposes programs interpreted by a universal Turing machine to produce byte sequences…
Harness-Zero: Harness Distillation via Agent-as-Harness
Harness-Zero uses a domain- or instance-optimized harness as training-time guidance, then transfers its behavior into model weights so the specialized harness can be removed at deployment. An agent-as-harness corrects…
VICBench: A Multi-Language Benchmark for Code Vulnerability Detection
The paper introduces VICBench, a new benchmark dataset for evaluating code vulnerability detection tools. VICBench contains 100 verified vulnerability-inducing commits (VICs) across 88 projects in Python, Java, and…
Rethinking Factor Sharing in Federated LoRA: A Rank-Aware Adaptive Approach
Researchers propose a new method for fine-tuning large language models in a federated learning paradigm using Low-rank adaptation (LoRA). They investigate two strategies, Share-A/Local-B and Share-B/Local-A, and reveal…
Improving the Realism of Synthetic Clinical Benchmarks Under Utility Constraints
The authors propose a method to improve the realism of synthetic clinical benchmarks for enterprise AI agents while maintaining their utility. They focus on healthcare settings where operational data is hard to access…
EgoGenesis: Egocentric World-Action Modeling with Online Anchored Projective Memory and Action-3D RoPE
The research presents a new method, EgoGenesis, for generating high-quality manipulation videos in an egocentric world for training embodied AI. The method uses a pretrained video generation prior and introduces two…
EgoGenesis: Egocentric World-Action Modeling with Online Anchored Projective Memory and Action-3D RoPE
The authors present a method called EgoGenesis for creating synthetic manipulation videos for embodied AI using egocentric video. The method uses a pretrained video generation prior and introduces two geometry-aware…
ScratchSim: A Procedural Synthetic Data Pipeline for Surface Scratch Detection
The paper introduces ScratchSim, a procedural rendering pipeline that generates large-scale synthetic data for surface scratch detection in industrial quality control. This pipeline uses BlenderProc and produces…
Synthetic data generation framework for quality control automation in gravure printing
Researchers have developed a synthetic data generation framework to help automate the quality control process in gravure printing. The framework generates high-fidelity images of common printing defects and uses them…
OmniaBench: Benchmarking General AI Agents Across Diverse Scenarios
A research paper titled 'OmniaBench' introduces a new benchmark for evaluating the capabilities of general AI agents across various scenarios, using diverse tasks drawn from app stores, product documents, industry…
A Multimodal Dataset for Large Language Model Applications in the Energy Domain
This paper introduces a new open-access dataset called mAIEnergy, which is designed to support the development of Large Language Models (LLMs) in the energy sector. The dataset contains various types of data including…
A Multimodal Dataset for Large Language Model Applications in the Energy Domain
The paper introduces a new open-access dataset called mAIEnergy which is a multimodal corpus designed to support Large Language Model (LLM) applications in the energy sector. It contains various types of data including…
Active Offline-to-Online Reinforcement Learning
Researchers propose an active policy selection method for fine-tuning in offline-to-online reinforcement learning (O2O-RL) settings, where the goal is to allocate a limited online interaction budget effectively among…
Localized LoRA-MoE: Block-wise Low-Rank Experts With Adaptive Routing
In this research, the authors propose Localized LoRA-MoE, a framework that combines localized spatial blocking with dynamic, context-conditioned routing for Large Language Models (LLMs) and high-dimensional perception…
JL1-CC&QA: Extending the JL1-CD Benchmark with Change Captioning and Question Answering
Researchers introduce JL1-CC&QA, a new benchmark for remote sensing that extends the JL1-CD dataset with two annotation layers: change captioning (CC) and change question answering (QA). The benchmark includes 5,000…
TRACE: Temporal Relationship-Aware Conversational Entrainment Detection in Dyadic Speech
Researchers have introduced a new dataset called DyadEE for studying emotional entrainment in two-person conversations. They also developed a model called TRACE that uses acoustic embeddings from emotion-fine-tuned…
TraceLab: Characterizing Coding Agent Workloads for LLM Serving
Researchers present TraceLab, a dataset of 4,300 coding-agent sessions collected from their own usage of Claude Code and Codex. The data reveals characteristics of these workloads, such as long autonomous loops, long…
Translating Natural Language to Strategic Temporal Specifications via LLMs
Researchers have developed a framework for translating natural language descriptions of strategic requirements into well-formed formulas using Large Language Models (LLMs). They created and curated a novel dataset for…
VLK: Learning Humanoid Loco-Manipulation from Synthetic Interactions in Reconstructed Scenes
Researchers present a method to learn humanoid locomotion and manipulation tasks by synthetically generating vision-language-kinematics (VLK) supervision in reconstructed scenes. They produce 48,000 paired trajectories…
Reliable Evaluation of MRI Motion Correction: Dataset and Insights
The paper discusses the difficulty in evaluating methods for correcting motion artifacts in medical imaging due to the lack of ground-truth data. To address this issue, they introduce PMoC3D, a dataset of real-world…
A Pipeline for Generating Longitudinal Synthetic Clinical Notes Using Large Language Models
A new pipeline for creating synthetic clinical notes is introduced to help develop AI systems in healthcare without compromising privacy. The pipeline generates structured, semi-structured, and unstructured clinical…
Inherited Circuits, Learned Semantics: How Fine-Tuning Creates Evasion Vulnerabilities Invisible to Standard Evaluation
The study shows that Language Models (LLMs) fine-tuned for security classification may overlook vulnerabilities introduced during fine-tuning itself. These models can learn token-level indicator semantics that preserve…
Inherited Circuits, Learned Semantics: How Fine-Tuning Creates Evasion Vulnerabilities Invisible to Standard Evaluation
A study shows that Language Models (LLMs) fine-tuned for security classification can miss vulnerabilities due to learning token-level indicator semantics that preserve accuracy on their training data but fail under…
WinDOM: Self-Family Distillation for Small-Model GUI Grounding
Researchers present WinDOM, a large dataset of bounding box training data for small GUI-grounding agents, which can be used for on-device deployment, accessibility tooling, and low-cost iteration. They also introduce…
