Fine-tuning

What it is

During fine-tuning, the model consumes a labeled dataset of input-output pairs to update its internal parameters via backpropagation. Unlike pretraining, which requires massive compute clusters, fine-tuning often requires only a few hundred to a few thousand examples and can be completed in minutes or hours on a single GPU. Techniques like LoRA allow you to freeze most of the model weights and train only a tiny adapter layer, drastically reducing the required memory and storage. This results in a new model instance specialized for your target behavior.

Why it matters

Fine-tuning matters when general-purpose models fail to follow strict output formats or lack domain-specific jargon required by your application. Relying solely on in-context learning through a massive system prompt can hit token limits, increase latency, and balloon your per-request cost. By fine-tuning, you can use a smaller, faster base model that achieves the same quality as a larger one. Choosing this path commits your team to managing model artifacts and deployment pipelines rather than just calling an external API.

In practice

You typically prepare a JSONL file containing your training examples and pass it to a training API or a library like Hugging Face's TRL. You will then monitor loss curves to ensure the model learns your task without drifting away from its original capabilities. In production, you deploy the resulting model adapter alongside the base model and select it via a model ID parameter in your request.

The tradeoff

The common trap is overfitting where the model masters your training data but loses the ability to generalize or follow basic instructions.

Where it appears

Research summaries that use Fine-tuning, each linked to its source paper.

• Selecting Better Training Data for Agents
  SWE-Prime: Fewer Trajectories, Better Performance

  The researchers introduced SWE-Prime, a method that selects a small, high-quality subset of training trajectories to improve software engineering agent performance.

• Automated Video Editing Through Verifier-Grounded Planning
  Plans You Can Check: Verifier-Grounded Learning of an Open-Weight Planner for Executable Video-Editing

  The authors introduce a framework for video-editing agents to generate and verify executable edit plans using a self-improving training loop.

• Using Coding Agents as World Brains
  Code World Model: Coding Agent as World Brain

  The paper introduces a framework where a coding agent generates deterministic code to manage world state, which then guides a video model to maintain visual consistency in simulated environments.

• Improving AI Code Generation With Robust Testing
  Robust Code RL via Faulty-Code-Driven Test case Synthesis and Dense Reward Shaping

  The paper introduces a framework called RobustTests that improves AI code generation by synthesizing diverse, failure-inducing test cases to guide reinforcement learning.

• Scaling Web Agents With Browser Sandboxes
  BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes

  Researchers built a large-scale web navigation dataset by orchestrating parallel browser sandboxes to generate diverse, high-quality interaction trajectories.

• Fixing Safety Losses During Model Reasoning
  Mitigating Reasoning-Induced Misalignment via Safety-Direction Penalty

  The researchers introduce a penalty method to prevent language models from sacrificing safety protocols when they are fine-tuned for improved reasoning tasks.

• Scaling Cyber Security Skills With AI
  CyberFactory: Scaling Cyber Security Capabilities with Instances from the Wild

  The researchers developed CyberFactory, a framework that leverages existing vulnerability data to train an AI model, OpenAegis, to improve security analysis performance.

• Improving Stable Critic Training for LLMs
  How to Train a Critic Stably and Efficiently

  The paper introduces BPCO, a method that stabilizes critic-based reinforcement learning, improving performance across various model sizes and tasks.

• Building Industrial Datasets from Technical Reports
  Industrial-Instruction: An End-to-End Framework for Building Instruction-Tuning and Benchmark Datasets from Industrial Technical Reports

  The authors present an end-to-end framework to automatically generate instruction-tuning and benchmark datasets from complex industrial technical documents.

• Securing AI Agents Using On-Policy Distillation
  SecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation

  SecOPD improves AI agent security against adaptive prompt injection by using on-policy distillation to provide fine-grained training signals that distinguish between trusted instructions and malicious data.

• Automated Environment Synthesis for Agent Training
  AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale

  AgentMercury automates the creation of executable business environments to improve agent performance and training scalability.

• Teaching Large Models to Memorize Documents
  Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization

  The researchers developed a staged training method called IAR to improve how models store and answer questions about specific document sets without needing retrieval systems.

• Controlling LLM Reasoning Costs via Self-Reflection
  Training-Free Inference-Time Self-Reflection and Cost-Bounded Early Stopping for Large Language Models

  The paper introduces a training-free inference-time protocol that uses a self-critique loop and a confirmed sentinel to improve reasoning accuracy while early-stopping redundant computations.

• Enabling Real-Time Interaction in Vision Models
  MOSS-VL Technical Report

  MOSS-VL introduces an architecture and training curriculum that allows vision-language models to process incoming video frames and generate responses simultaneously.

• Learning Full 3D Objects from LiDAR
  GhostPoint: Self-Supervised Representation Learning by Hallucinating Occluded LiDAR Structure

  GhostPoint improves 3D object detection by training models to predict the hidden structure of objects that are partially occluded in LiDAR sensor data.

• Improving Robot Reaction Times for Manipulation
  Reflex: Enabling Fast and Predictive Vision-Language-Action Models for Reaction-Critical Manipulation

  The paper introduces ReflexVLA, a vision-language-action model architecture that uses future prediction and optimized inference to improve performance in time-sensitive robotics tasks.

• Evaluating Foundation Models for Movement Data
  Foundation models for movement data: Are they ready for prime-time?

  The paper benchmarks four foundation models against traditional supervised learning baselines across 19 movement and health monitoring tasks to determine their practical effectiveness.

• Simulating Realistic Fog for Improved Perception
  A Data Efficiency Study of Synthetic Fog for Object Detection Using the Clear2Fog Pipeline

  The Clear2Fog pipeline enables generation of physically realistic fog across RGB and LiDAR modalities to improve the accuracy of autonomous vehicle perception systems.

• Reprogramming Models into Assertive Socratic Assistants
  Behavioral Reprogramming of Open-Weights Models: Cognitive Plasticity and Alignment Bounds

  The paper demonstrates how to fine-tune open-weights models to shift from passive assistant behaviors to a proactive Socratic persona using targeted parameter-efficient techniques.

• Running Virtual Agents on Edge Hardware
  Enhancing Virtual Agents through SLMs and Edge-Computing: An Exploratory Evaluation of Think and Memory Processes

  The paper explores using Small Language Models on NVIDIA Jetson Orin NX hardware to handle cognitive tasks like service routing and memory management for virtual agents.

• Improving Driving Video Search via Motion
  TraVEL: Trajectory-Guided Video Embedding Learning for Driving-Video Retrieval

  TraVEL improves driving video retrieval by training embedding models to prioritize ego-vehicle movement patterns over static visual shortcuts.

• Improving Web Agent Performance via Co-Synthesis
  SynWeaver: Website-Prior Task and Trajectory Co-Synthesis for Web Agents

  SynWeaver improves web agent accuracy by co-synthesizing website-specific tasks and execution trajectories to overcome the lack of supervision on unseen websites.

• Stopping LLM Safety Bypass Attacks
  Refusing Intent, Not Form: Wrapper-Based Intent-Group Supervision for LLM Safety

  The paper introduces a fine-tuning method to prevent models from being tricked by malicious prompt wrappers that bypass safety filters or cause over-refusal of benign tasks.

• Improving Embodied Agents Without Retraining
  Self-Evolving Embodied Agents via Skill-Harness Evolution

  The SHAPER method improves agent performance in new environments by evolving textual skills and harnesses while keeping the underlying model parameters frozen.

• Improving AI Reasoning via Automated Harnesses
  AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses

  Researchers developed a method where a strong builder model automatically generates and refines scaffolding logic to improve the performance of smaller language models on reasoning tasks.

• Improving Burmese Clinical Speech Recognition
  myMediWhisper: Construction of Burmese Medical Speech Corpus and Whisper Fine-Tuning for Clinical Dialogue ASR

  The researchers constructed a Burmese medical speech corpus and fine-tuned Whisper models to improve automatic speech recognition for clinical dialogues.

• Improving Surgical Robot Learning with Video
  Surgical WAM: A World-Action Model for Data-Efficient Surgical Robot Learning

  The researchers developed a World-Action Model that leverages action-free video pretraining to significantly boost the performance of surgical robots when labeled demonstration data is limited.

• Extracting Structured Knowledge from Language Models
  REAP: Relation-Aware Elicitation and Parsing for Closed-Book Knowledge Base Construction from LLMs

  The REAP system improves closed-book knowledge base construction by using a two-stage process of relation-aware prompting and hybrid parsing to extract structured data from large language models.

• Efficient Multilingual Routing for Short Texts
  A Cost-Efficient Routing Pipeline for Multilingual Short-Text Classification Using Small Language Models

  The paper introduces a cost-aware routing strategy that selects between direct multilingual processing and translation-based English classification to optimize performance for weaker languages.

• Removing Specific Concepts from Diffusion Models
  PEAK: Precise and Persistent Concept Erasure via k-Sparse Autoencoders

  PEAK uses sparse autoencoders to precisely identify and suppress target concepts in diffusion models while maintaining overall generation quality.

• Tracing Model Misalignment to Specific Data
  Data Attribution of Emergent Misalignment with Persona Features

  Researchers identified specific pre-training documents that cause emergent misalignment in language models and demonstrated that synthetic instruction tuning exacerbates this behavior.

• Adapting Robot Vision to New Camera Views
  AnyCamVLA: Zero-Shot Camera Adaptation for Viewpoint Robust Vision-Language-Action Models

  AnyCamVLA improves robot task performance in new camera environments by synthesizing training-viewpoint images in real-time before processing them with a pre-trained policy.

• Optimizing LLMs for Better Performance and Efficiency
  Can We Optimize the Performance-Carbon Emission Break-Even Point?: The Quest for Greener LLMs

  The paper introduces a joint loss optimization technique that fine-tunes LLMs to improve task accuracy while simultaneously minimizing inference carbon emissions.

• Improving Multimodal Models Without External Labels
  Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spots

  The researchers developed a self-distillation technique for multimodal large language models that sharpens visual perception by identifying and training on internal counterfactual blind spots.

• Improving Visual Tool Use for Models
  OpenVisTool: An Open Recipe for Synthesizing Instructive Visual Tool-Use Trajectories

  OpenVisTool introduces a training method that teaches models to use external visual tools only when necessary, improving performance over fixed image encoding.

• Improving Navigation with 3D Scene Awareness
  WNM-3D: A World Navigation Model with 3D Scene Conditioning for Closed-Loop VLN

  WNM-3D introduces a geometry-aware conditioning mechanism to improve closed-loop vision-language navigation using a diffusion transformer.

• Improving Video Captioning with Atomic Rewards
  AVCap: Reinforcing Audio-Video Joint Caption with Detail-Aware Reward

  The paper introduces a new framework for training video captioning models that uses fine-grained, atomic-level rewards to improve accuracy in describing video content.

• Reading ESG Concept Content from LLMs
  Measuring Concept Content in Text from LLM Activations: ESG Evidence from Concept Vectors and Linear Probes

  The paper demonstrates that monitoring internal activations of frozen LLMs can measure ESG concept content in text as an alternative to fine-tuning.

• Distilling Hierarchical Memory for Agent Models
  Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory

  The researchers developed a training-free framework that distills complex strategies and tool-calling logic from large teacher models into smaller, more efficient student models.

• Decoupling Agent Scaffolding to Improve Planning
  DCAS: Decoupling CLI Agent Scaffolding to Internalize Planning across Scaffolds

  The researchers introduced DCAS, an interception layer that decouples agent planning from specific CLI environments to enable cross-environment training and better performance.

• Improving Mobile GUI Agents with Hindsight
  The Next Screenshot Knows: Gated Hindsight Distillation for Mobile GUI Agents

  The researchers introduce Gated Hindsight Distillation to help GUI agents learn from future screenshots when current screen data is insufficient for decision making.

• Optimizing RAG Latency via Cached Information Nuggets
  CoinRAG: Contextualized Information Nugget KV Cache Reuse for Long-Context RAG

  CoinRAG reduces RAG latency and computational redundancy by precomputing and reusing specific information nugget representations within the model KV cache.

• Automated Patient Positioning for Radiography
  Patient Pose Assessment Using a CT-Based Framework for Synthetic Data Generation

  The paper introduces a synthetic data generation framework that allows AI to accurately assess patient poses for X-rays by training on generated depth images and radiographs.

• Learning Compositional Analysis of Visual Media
  Learning visual representations for compositional analysis of artworks and photographs

  The paper introduces a human-inspired pipeline using Object-Centric Learning and graph networks to model image composition more effectively than frozen foundation models.

• Learning 3D Robot Actions from Video
  LAWM-3D: Learning 3D-Aware Latent Actions from Human Videos for Generalizable Robot World Models

  LAWM-3D enables robots to learn 3D-aware actions by training world models on human videos using a new geometric alignment method.

• Improving Medical Time Series Classification
  Is Self-Pretraining really useful to improve diagnosis in medical Time Series?

  The researchers applied self-pretraining to transformer models to boost accuracy in medical time series classification without requiring external data.

• Optimizing Small Models for Human Behavior
  Small Foundation Models of Human Cognition and Behaviour

  The paper tests whether small language models fine-tuned on behavioral data use structural reasoning or statistical shortcuts to predict human task performance.

• Teaching Models to Ignore Misleading Signals
  Learning When to Trust via Selective Context Preference Optimization

  The paper introduces the SCOPE framework to prevent language models from abandoning correct answers when they encounter deceptive or irrelevant context.

• Retrieving Better Insights for AI Agents
  InsightEmb: Learning Action-Intent Embeddings for Agentic Insight Retrieval

  The paper introduces InsightEmb, a method that improves agentic tasks by matching an agent's current state to helpful heuristic insights rather than relying on standard semantic similarity.

• Benchmarking Global Spatial Awareness in VLMs
  GST-Bench: Can VLMs Develop Global Spatial Awareness from Video?

  The authors introduce GST-Bench to evaluate and improve how vision-language models maintain consistent spatial understanding across long, continuous video streams.

• Improving Temporal Reasoning in Multimodal Models
  ChronoVision: Temporal Reasoning via Latent State Reconstruction

  ChronoVision introduces a visual-focused training framework to help multimodal large language models track and reason about continuous changes in images.

• Adapting Tabular Foundation Models Without Retraining
  SkillTFM: Gated Skill Evolution for Training-Free Adaptation of Tabular Foundation Models

  SkillTFM enables tabular foundation models to adapt to new tasks and data distributions by dynamically retrieving and evolving skills from a pre-verified skill bank.

• Argus Agentic Runtime for Long Tasks
  Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning

  Argus is a persistent runtime system that improves research agent performance by evolving operational state and project objectives alongside human guidance.

• Teaching Logic Before Language to LLMs
  Logic Before Language: Pre-pretraining on Formal Derivations Fosters Skill Acquisition and Compressibility

  Researchers improved language model learning efficiency by pre-pretraining a Transformer backbone on formal logic derivation sequences before standard language training.

• Evaluating Agent Self-Evolution on Business Tasks
  GDPevo: Evaluating Agent Self-Evolution on Real Business Tasks

  The authors introduce GDPevo, a benchmark and automated pipeline designed to evaluate and improve how agents evolve their performance on complex enterprise workflows.

• Connecting Vision Language Models To Robots
  RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control

  The paper introduces vision language action models, which incorporate internet scale web data directly into robotic control to improve generalization and semantic reasoning.

• Building a General Purpose Visual Assistant
  Visual Instruction Tuning

  The paper introduces LLaVA, a multimodal visual assistant built by connecting a visual encoder to a language model and fine-tuning them on automatically generated instruction data.

• Evaluating Large Language Models Trained on Code
  Evaluating Large Language Models Trained on Code

  The paper investigates the program synthesis and code-writing capabilities of large language models trained on code, focusing on generating standalone Python functions from docstrings and measuring functional correctness.

• GPT-3 Few-Shot Language Learning Performance
  Language Models are Few-Shot Learners

  The paper introduces a 175-billion parameter model capable of performing tasks with zero or few examples provided in the prompt without needing model weight updates.

• Training Deep Bidirectional Language Models
  BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

  This paper introduces BERT to enable pre-training of deep bidirectional representations by jointly conditioning on both left and right context in all layers.

• Universal Language Model Fine-Tuning
  Universal Language Model Fine-tuning for Text Classification

  The paper introduces ULMFiT, a universal transfer learning method that significantly outperforms existing approaches across six text classification tasks.

• Using Transformers for Image Recognition
  An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

  The paper demonstrates that a pure Transformer architecture, applied directly to sequences of image patches, can achieve excellent image classification results compared to state-of-the-art convolutional networks while requiring substantially fewer computational resources to train.

• Context Aware Image Similarity Metrics
  The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric

  The paper introduces TPIPS, a text-prompted metric that allows developers to calculate image similarity based on specific visual attributes rather than using generic, aspect-agnostic scores.

• Automated Video Editing Through Flow Mimicry
  FlowMimic: Mask-free Visual Editing and Generation with Pixel-pair Warped Flow Field for Online Video Editing Data Generation and Modality Mimicry

  FlowMimic generates synthetic video editing data in real time by applying temporal flow fields to existing image editing samples, removing the need for labor-intensive mask annotations.

• Efficient Data Selection for LLM Training
  PPL-Factory: Task-Aware and Budget-Aware Data Selection from Language Modeling to Reasoning

  PPL-Factory is a method that selects the most informative subset of training data to improve fine-tuning performance while reducing computational overhead.

• Efficient Robotic Control via Patch Representations
  Patch Policy: Efficient Embodied Control via Dense Visual Representations

  The paper introduces a method that replaces compressed global visual vectors with uncompressed patch tokens to improve robotic manipulation precision while maintaining high computational efficiency.

• Improving Assamese Speech Recognition Performance
  Robust Assamese Speech Recognition through Controlled Fine-Tuning of Whisper Models

  Researchers improved Assamese speech recognition by fine-tuning the Whisper model using a combined dataset of validated and crowd-sourced audio.

• Xiaomi-Robotics-1: Scaling Robot Action Models
  Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

  Xiaomi-Robotics-1 introduces a large-scale vision-language-action model and a scalable auto-labeling pipeline to overcome data bottlenecks in robotics, achieving strong performance on robot benchmarks.

• Training Large Context Models On Low VRAM
  Long-Context Fine-Tuning with Limited VRAM

  This paper introduces a memory-efficient training technique called Hierarchical Global Attention to process significantly longer token sequences on constrained GPU hardware.

• Benchmarking Visual Evidence in Issue Localization
  MM-IssueLoc: A Controlled Benchmark for Evaluating Visual Evidence in Multimodal Repository-Level Issue Localization

  The researchers developed the MM-IssueLoc benchmark to evaluate how visual evidence like screenshots affects AI-driven repository-level issue localization.

• Converting Document Images to Markdown Efficiently
  OvisOCR2 Technical Report

  OvisOCR2 is a model designed to parse visually rich documents into structured Markdown in a single pass.

• Optimizing OCR for Multiple Historical Scripts
  Multi-Expert Routing for Multi-Domain Low-Resource OCR: A Manchu Case Study

  The authors implement a domain routing system that dispatches page images to specialized OCR models, enabling accurate text extraction across varied historical Manchu writing styles.

• Agentic Framework for Traffic Anomaly Understanding
  TAU-Agent: An Agentic Retrieval-Augmented Framework for Traffic Anomaly Understanding

  TAU-Agent is an agentic framework designed to improve traffic anomaly detection by using retrieval-augmented generation to integrate video descriptions and object trajectories.

• Teaching LLMs Clinical Reasoning for ICUs
  Teaching LLMs How ICU Physicians Approach Clinical Reasoning Through OMOP-Aligned Retrieval Improves Reasoning Across Clinical Domains

  The authors created the ICU-REACT dataset and a corresponding family of fine-tuned models to improve LLM performance in identifying and reasoning over patient data for critical care.

• Automated Terminal Task Synthesis Framework
  FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis

  The paper introduces FACET, a framework for generating consistent, executable terminal tasks by grounding task artifacts like instructions and verifiers in a shared containerized state.

• Expanding Creative Writing Beyond Simple Stories
  Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion

  Researchers developed a method using attribute guided genre expansion to train language models on diverse creative formats beyond basic narrative generation.

• Efficient Text Guided Image Upscaling
  MagnifiQ: Patch-aware Text Guided Progressive Upscaling for High-Resolution Image Restoration

  MagnifiQ uses a modular patching architecture and LLM-based text prompts to perform efficient high resolution image restoration.

• Evaluating Detection of AI-Generated Crisis Videos
  Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination

  Researchers built a new benchmark, RA-Bench, to systematically test how well current detection methods identify AI-generated videos during real-world social crises.

• Structuring Scientific Knowledge from Full Text
  MUSE: A Full-Text Cross-Domain Knowledge Base of Scientific Problems, Solutions, and Rationales

  The paper introduces MUSE, a large-scale knowledge base of 36,960 structured problem-solution-rationale triplets extracted directly from full-text scientific papers.

• Evaluating Language Models on Real Software Issues
  SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

  The paper introduces SWE-bench, a new benchmark that tests language models on resolving real-world GitHub issues by navigating large repositories and executing unit tests.

• Teaching Language Models to Use APIs
  Toolformer: Language Models Can Teach Themselves to Use Tools

  The researchers developed a self-supervised method allowing language models to learn how and when to invoke external APIs to overcome limitations in calculation and factual retrieval.

• Constitutional AI for Harmless Assistants
  Constitutional AI: Harmlessness from AI Feedback

  The paper introduces Constitutional AI to train helpful and harmless AI assistants using self-critique and AI feedback instead of relying solely on manual human oversight.

• Flamingo Visual Language Model
  Flamingo: a Visual Language Model for Few-Shot Learning

  Flamingo is a visual language model that adapts to novel multimodal tasks using only a handful of annotated examples.

• Synthetic Training Data for API Agents
  Environment-free Synthetic Data Generation for API-Calling Agents

  The paper introduces a synthetic data generation framework that enables training API-calling agents without needing fully operational backend environments.

• Steering LLM Psychotherapy Interactions
  Move by Move: Measuring and Steering How LLMs Conduct Psychotherapy

  The authors introduce a method to steer LLM behavior in therapy sessions by exposing a clinical move ontology as tools, which improves alignment with human therapists.

• Robust Speech Recognition via Large Scale Weak Supervision
  Robust Speech Recognition via Large-Scale Weak Supervision

  The paper develops a robust speech recognition system that works reliably out of the box across a broad range of environments, tasks, and languages without requiring supervised fine tuning of a decoder for every deployment distribution.

• Selecting Better Training Data for Agents
  SWE-Prime: Fewer Trajectories, Better Performance

  The researchers introduced SWE-Prime, a method that selects a small, high-quality subset of training trajectories to improve software engineering agent performance.

• Removing User Interface Overlays from Gameplay
  Game2World Engine: Unlocking In-the-Wild Gameplay Videos for World Model Training

  The authors introduce GameCleaner and the Game2World engine to remove distracting interface elements from game footage to improve the training of world models.

• Improving AI Code Generation With Robust Testing
  Robust Code RL via Faulty-Code-Driven Test case Synthesis and Dense Reward Shaping

  The paper introduces a framework called RobustTests that improves AI code generation by synthesizing diverse, failure-inducing test cases to guide reinforcement learning.

• Improving AI Reasoning Through Verifiable Distillation
  On-policy Distillation with Verifiable Reward

  The paper introduces a method called OPDVR that aligns reinforcement learning signal with task success during model distillation to improve reasoning performance.

• Automated Self-Improvement for LLM Judges
  RecurSE: Bounded Recursive Self-Evaluation for LLM Rubric Judges

  RecurSE enables LLM-based judges to improve their evaluation performance by creating a bounded, self-correcting feedback loop that eliminates the need for external gold standard rewards.

• Improving Stable Critic Training for LLMs
  How to Train a Critic Stably and Efficiently

  The paper introduces BPCO, a method that stabilizes critic-based reinforcement learning, improving performance across various model sizes and tasks.

• Building Industrial Datasets from Technical Reports
  Industrial-Instruction: An End-to-End Framework for Building Instruction-Tuning and Benchmark Datasets from Industrial Technical Reports

  The authors present an end-to-end framework to automatically generate instruction-tuning and benchmark datasets from complex industrial technical documents.

• Training Adaptable Agents for Live Streaming
  Training Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical Report

  The paper introduces a method called Harness Evolution that improves agent adaptability and performance for live-streaming environments by decoupling execution settings from the base model.

• Optimizing MedSAM Fine-Tuning for Robustness
  When Adaptation Hurts: Connecting Representational Drift to OOD Failures in MedSAM Fine-Tuning

  The paper identifies that representational drift in the decoder and output layers significantly impacts out of distribution performance when fine-tuning the MedSAM medical image segmentation foundation model.

• Recovering Compressed 4 Bit LLMs
  Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs

  The paper introduces Quantization-Aware Healing, a practical recipe for recovering compressed 4-bit large language models, and uses it to produce the open-weight model Hypernova-60B.

• Improving Robot Policies Without Retraining
  Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning

  The paper presents a method that enables robot policies to self-improve through iterative deployment without the need to modify the original policy weights.

• Improving Efficiency and Expressivity in TTT
  Rethinking Expressivity and Efficiency in Test-Time Training

  The paper introduces E2-TTT, a new method for Test-Time Training that uses chunk-wise updates to achieve higher performance while maintaining computational efficiency.

• Selective Safety Routing for Language Models
  CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment

  The paper introduces a routing mechanism that applies safety interventions only when harmful inputs are detected, preserving model utility for benign prompts.

• Improving AI Model Preference Optimization
  Manifold Drift in Flow Preference Optimization: A Root Cause of Reward Hacking

  The paper introduces ThermoDPO, a new training method that stabilizes generative model output by preventing reward-driven distortion of the underlying data distribution.

• Efficient Hyperparameter Optimization for Large Models
  Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts

  The paper introduces a method to extrapolate optimal learning rates for Mixture of Experts models using small-scale proxy runs to avoid expensive full-scale sweeps.

• Improving AI Tool Use Through Mid-Training
  MidTool: Mid-training Data Synthesis for Agentic Tool Use

  The researchers created a 20.3B-token corpus called MidTool-Mix to improve agentic tool-use capabilities in models during the mid-training phase rather than relying solely on post-training.

• Teaching Large Models to Memorize Documents
  Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization

  The researchers developed a staged training method called IAR to improve how models store and answer questions about specific document sets without needing retrieval systems.

• Fine-Tuning LLM Agents with Less Memory
  Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements

  The paper introduces Agentic ESOpt, a method for fine-tuning LLM agents that replaces traditional backpropagation with population-based parameter perturbations to reduce memory requirements during training.

• Expanding Creative Writing Beyond Simple Stories
  Scaling Creative Writing Beyond Story-Centric Data with Attribute-Guided Genre Expansion

  Researchers developed a method using attribute guided genre expansion to train language models on diverse creative formats beyond basic narrative generation.

• Processing Recipe Data with Transformers
  RecipeNet: A Hierarchical Transformer for Recipe Data

  RecipeNet is a hierarchical transformer model designed to process heterogeneous recipe data with variable schemas and sequential procedural steps.

• Optimizing Neural Network Model Merging
  Multi-Objective Bayesian Optimization for Model Merging

  The paper introduces a multi-objective optimization framework to automatically select the best merge parameters for combining specialized neural network models.

• Learning Full 3D Objects from LiDAR
  GhostPoint: Self-Supervised Representation Learning by Hallucinating Occluded LiDAR Structure

  GhostPoint improves 3D object detection by training models to predict the hidden structure of objects that are partially occluded in LiDAR sensor data.

• Automating Lean 4 Proof Formalization
  MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement

  MathForm improves mathematical autoformalization by integrating external retrieval and verification-guided feedback loops to generate verified Lean 4 code.

• Personalizing Images Without Paired Data
  CRAFT: Constrained Reward via Attention Fine-Tuning for Subject Personalization without Composed Targets

  CRAFT enables subject-driven image personalization using only reference-side supervision, completely eliminating the need for costly multi-stage curation pipelines that generate paired reference and composed target data.

• Optimizing Data Repetition in LLM Pretraining
  Scaling Domain Data Repetition in LLM Pretraining

  The paper examines how to effectively repeat high-quality domain data during the pretraining phase as model sizes and training token budgets scale.

• Optimizing Vision Language Model Training Efficiency
  Rollplex: Cross-Phase GPU Spatial Sharing for Vision Language Model Post-Training

  Rollplex increases GPU utilization during vision-language model post-training by overlapping prompt processing with rollout decoding to eliminate serial execution bottlenecks.

• Transferring Reasoning Capabilities Between Language Models
  SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning

  The paper introduces a method called SimpleOPD that distills advanced mathematical reasoning from a high-performance teacher model into smaller student models across different architectures.

• Improving AI Vision Without External Data
  Self-Supervised Visual On-Policy Distillation

  The paper introduces a self-supervised method to improve vision model performance by distilling knowledge from a student to an EMA teacher without needing ground-truth labels or extra rewards.

• Optimizing Muon for Low Rank Adapters
  Approximate Muon with low-rank adapters

  The paper introduces sMuon, an algorithm that enables the use of the Muon optimizer within low rank adaptation techniques to improve fine-tuning performance.

• Foundation Model for Infrared Chemical Sensing
  Simulation-to-real transfer learning for infrared spectroscopic chemical sensing and analysis from molecules to complex samples

  UltraIR is a foundation model that uses simulation-to-real transfer learning to improve the accuracy and scalability of infrared spectroscopy analysis.

• Unifying Cardiac Signals With AI
  CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation

  The paper introduces CardioState-JEPA, a unified cardiac foundation model that learns a single shared representation across heterogeneous signals like electrocardiography, photoplethysmography, and phonocardiography by accounting for physiological delays.

• Evaluating Foundation Models for Movement Data
  Foundation models for movement data: Are they ready for prime-time?

  The paper benchmarks four foundation models against traditional supervised learning baselines across 19 movement and health monitoring tasks to determine their practical effectiveness.

• Reprogramming Models into Assertive Socratic Assistants
  Behavioral Reprogramming of Open-Weights Models: Cognitive Plasticity and Alignment Bounds

  The paper demonstrates how to fine-tune open-weights models to shift from passive assistant behaviors to a proactive Socratic persona using targeted parameter-efficient techniques.

• Explaining Neural Network Learning Dynamics
  Neural Quadratic Forms: A Unified Minimal Model for Sudden Learning and Scaling Laws

  The authors derive a mathematical framework that models how neural network training converges using a small set of variables regardless of the model size.

• Controlling Knowledge Acquisition in Language Models
  LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure

  The researchers developed a pedagogically constrained training environment to build models that learn only within a specific academic scope.

• Measuring Data Influence in LLM Training
  Measuring Task-Agnostic Training Data Influence Across Language Model Pretraining

  The researchers developed a task-agnostic method to quantify how individual training examples affect a language model's final parameters without needing to retrain the model.

• Improving Model Distillation via Task Sensitivity
  CROP: Task Relevance via Counterfactuals for Selective On-Policy Distillation

  The paper introduces a method called CROP to selectively focus model distillation on task-relevant information by measuring sensitivity to counterfactual prompts.

• Making AI Models Learn Their Own Context
  Latent On-Policy Self-Distillation

  The researchers developed a method to replace hand-designed improvement rules with a system that learns to generate its own contextual guidance for model training.

• Stabilizing Cascaded Image Restoration and Detection
  Delving into Cascaded Instability: A Lipschitz Continuity View on Image Restoration and Object Detection Synergy

  The authors introduce Lipschitz-regularized object detection to fix instability caused by functional mismatches when chaining image restoration with detection models.

• Improving AI Labeling with Advanced Models
  CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers

  The paper introduces CW-BASS v2 to fix pseudo-labeling errors that occur when high-performance foundation models become overly confident and biased during training.

• Automated Malware Detection Model Updates
  Concept Drift Detection and Adaptive Retraining of Malware Classification Models

  The paper introduces drift-aware retraining strategies to maintain malware classification accuracy while minimizing the number of model updates compared to periodic retraining.

• Stopping LLM Safety Bypass Attacks
  Refusing Intent, Not Form: Wrapper-Based Intent-Group Supervision for LLM Safety

  The paper introduces a fine-tuning method to prevent models from being tricked by malicious prompt wrappers that bypass safety filters or cause over-refusal of benign tasks.

• How Instruction Tuning Affects Model Confidence
  Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity

  Researchers evaluated how instruction tuning influences the verbalized confidence and lexical diversity of rationales generated by three popular large language models.

• Embedding Assistant Personas During Model Training
  Synthetic Persona Pretraining: Alignment from Token Zero

  Researchers developed Synthetic Persona Pretraining to embed desired assistant behaviors into language models starting from the very first token of training.

• Building Capable Models With Only Permissible Data
  DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data

  Researchers developed Mimir v1, a foundation model trained on 161 permissible datasets to ensure compliance without sacrificing performance in English, Math, and Danish tasks.

• Improving Attention Gating for Language Models
  Hybrid Gated Attention

  The paper introduces Hybrid Gated Attention, a technique that improves training stability and performance in language models by modifying how attention mechanisms handle gating, matrix factorization, and head interactions.

• Understanding Massive Activations in Hybrid Large Language Models
  Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus

  The paper tracks massive activation tokens across hybrid linear attention large language models to understand how layerwise hybridization reshapes internal activation dynamics.

• Preference Tree Optimization for Dialogue Systems
  Preference Tree Optimization: Enhancing Goal-Oriented Dialogue with Look-Ahead Simulations

  The paper introduces Preference Tree Optimization, a method that uses look-ahead simulations to enhance goal-oriented dialogue systems by training models to prefer paths that lead to better future conversation outcomes.

• Reducing Object Hallucination in Multimodal Models
  Context Blindness in DPO: Mitigating Object Hallucination in MLLMs via Context-Calibrated Preference Optimization

  The paper introduces Context-Calibrated DPO to force models to better utilize contextual information and reduce object hallucinations.

• Long Contexts Weaken Model Internal Knowledge
  Information Abundance Paradox: Long-Context Training Undermines Parametric Knowledge

  The researchers found that training language models on long documents makes them rely more on provided text and less on their own internal knowledge, leading to worse performance when that context is missing.

• Improving Burmese Clinical Speech Recognition
  myMediWhisper: Construction of Burmese Medical Speech Corpus and Whisper Fine-Tuning for Clinical Dialogue ASR

  The researchers constructed a Burmese medical speech corpus and fine-tuned Whisper models to improve automatic speech recognition for clinical dialogues.

• Improving Surgical Robot Learning with Video
  Surgical WAM: A World-Action Model for Data-Efficient Surgical Robot Learning

  The researchers developed a World-Action Model that leverages action-free video pretraining to significantly boost the performance of surgical robots when labeled demonstration data is limited.

• Optimizing Memory for Recommender Training
  Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training

  The paper provides a selection rule to optimally allocate a fixed memory budget between batch size and negative samples when training large scale recommender systems.

• Structuring Scientific Knowledge from Full Text
  MUSE: A Full-Text Cross-Domain Knowledge Base of Scientific Problems, Solutions, and Rationales

  The paper introduces MUSE, a large-scale knowledge base of 36,960 structured problem-solution-rationale triplets extracted directly from full-text scientific papers.

• Tracing Model Misalignment to Specific Data
  Data Attribution of Emergent Misalignment with Persona Features

  Researchers identified specific pre-training documents that cause emergent misalignment in language models and demonstrated that synthetic instruction tuning exacerbates this behavior.

• Finding Hidden Failures in LLM Training
  SCOUT: Symmetric Consensus Outlier Detection for Failure Localization in LLM Pre-Training

  SCOUT is a system that identifies and localizes latent hardware or communication failures during distributed large language model pre-training by comparing the behavior of identical parallel processing units.

• Improving Image Generation Quality via Adversarial Training
  AdvFD: Boosting Visual Generation via Adversarial Fr'echet Distance Loss

  The researchers developed AdvFD, a new training method that prevents image generators from exploiting static metrics to inflate their performance scores without actually improving visual quality.

• Interleaving Visual Objects for Better Alignment
  MultiModal Code-Switching: Interleaving Visual Objects into Language for Explicit Object-Level Alignment

  MultiModal Code-Switching improves multimodal model alignment by replacing text tokens with visual object embeddings during pretraining.