Fine-tuning

Latest: EOS Tokens Disagree: Paper Traces Length Inflation in On-Policy Distillation to Termination Mismatch · arxiv.org, Sep 18, 2026

Thème Mots-clés suivis : fine-tuning, fine tuning, lora, rlhf

16 AI Weekly issues and 46 live stories from the last 30 days mention Fine-tuning.

33 mentions au total 3/20 numéros récents -100% vs référence pic 2/numéro suivi depuis 2026-04-18

Tendance des mentions — 20 derniers numéros

Actus en direct (46)

EOS Tokens Disagree: Paper Traces Length Inflation in On-Policy Distillation to Termination Mismatch

arxiv.org · 2026-09-18 · Score : 21

Shanghai AI Lab's SP3O Cures PPO 'Value Flattening' With Sparse Critic Supervision on Three States per Response

arxiv.org · 2026-09-17 · Score : 25

Cambridge's XConf Beats 10-Sample Self-Consistency on 23 of 24 AUROC by Recalling Past Reasoning Episodes

arxiv.org · 2026-09-17 · Score : 25

ImpossibleRubrics Benchmark Shows LLM-Generated Rubrics Get Gamed 8–26% as RL Reward Signals

huggingface.co · 2026-09-15 · Score : 17

Paper: On-Policy Reverse Distillation Lets Student LLMs Surpass Weak Teachers Faster Than Verifier-Only RL

arxiv.org · 2026-09-08 · Score : 19

IBM's DRACO Paper Redistributes Rubric Rewards Across Steps, Lifts AppWorld Agent Score 15.9 Points

arxiv.org · 2026-09-03 · Score : 21

'One Training Example' Paper Shows On-Policy Distillation Is Data-Overfed, Algorithm-Starved

huggingface.co · 2026-09-03 · Score : 21

Debias-SparseGPT Paper Adds Bias-Aware Pruning to LLM Compression, Retains Fairness at Sparse Weights

huggingface.co · 2026-09-03 · Score : 17

Cliff Paper: Zeroing In on the First Mistake in a Rollout Boosts RL-Trained LLMs 15% Over Distillation

huggingface.co · 2026-09-02 · Score : 23

Meta and Princeton Paper: Mid-Training KL Distillation Boosts Reasoning But Slows Factual Recall

arxiv.org · 2026-09-01 · Score : 18

DiagEvo Paper: Self-Evolution via Hierarchical Error Memory Beats R-Zero by 4.5 Points on Qwen3-8B Math

huggingface.co · 2026-09-02 · Score : 18

Paper: Split Qwen3-32B Into GRPO Experts, Merge via SLERP, Match Qwen3-235B on Enterprise Traffic

huggingface.co · 2026-09-02 · Score : 19

NoRA: Normalized LoRA Method Speeds Convergence and Cuts Catastrophic Forgetting at Zero Extra Cost

huggingface.co · 2026-09-01 · Score : 20

Trajectory Raises $40M Sequoia-Led Round at $300M for Continual-Learning Agents

theinformation.com · 2026-08-11 · Score : 21

ABSeeker Paper Turns Sparse Trajectory Outcomes Into Dense Step Rewards for Long-Horizon Search Agents

huggingface.co · 2026-08-05 · Score : 19

TurnSight Adds Turn-Level Hindsight Self-Distillation to Tool-Integrated Reasoning, Beats Trajectory-Level RL

huggingface.co · 2026-08-04 · Score : 17

Arxiv Paper Pins LLMs' Tabular Weakness on Dimensionality, Only Method Whose Accuracy Falls as Features Grow

arxiv.org · 2026-08-03 · Score : 18

BIGAI-PKU 'GradCuit' Adds Credit-Assigned Gradient Flow to Test-Time Latent Reasoning

huggingface.co · 2026-08-03 · Score : 20

DAPD Paper Blames 'Privilege Illusion' in On-Policy Distillation, Beats OPSD by 2+ Points Across Qwen3 Scales

huggingface.co · 2026-08-03 · Score : 20

CUHK-Lightspeed 'Skill-α' Uses Rollback Rewards to Learn Agent-Skill Edits, Beats SkillPro by 6.7 pts

huggingface.co · 2026-08-03 · Score : 15

β-OPSD Paper Generalizes On-Policy Self-Distillation Into a Tunable Family, Beats Vanilla OPSD on Math Reasoning

huggingface.co · 2026-07-31 · Score : 17

Tsinghua's Memory Decoder at Scale Pairs 410M Base With 6.9B External Memory, Matches Pythia-12B

huggingface.co · 2026-07-30 · Score : 22

Peking-Kling's Flux-OPD Uses Evolving Contexts as In-Training Supervision, Beats OPD Baselines on Open-Ended Tasks

huggingface.co · 2026-07-31 · Score : 20

CoRT Uses Counterfactual Replay to Add Token-Level Credit to Rubric-Based GRPO, Gains 4.4 Points on Average

huggingface.co · 2026-07-29 · Score : 15

Meta FAIR Ships DMC-Optim RL Framework, Lifts Qwen 2.5 7B Code-Optimization Pass Rate From 18% to 31%

arxiv.org · 2026-07-28 · Score : 21

Zhejiang and Alibaba's Relay-OPD Fixes On-Policy Distillation's Prefix Failure With Teacher Handoffs, Beats OPD 5.7% on Math

huggingface.co · 2026-07-29 · Score : 23

Paper Maps Multi-Turn Long-Horizon Planning Across Pre-Training, Post-Training and Multi-Teacher Distillation

huggingface.co · 2026-07-27 · Score : 17

MAPD Paper: Multi-Agent Protocol Distillation Beats SDAR by 3-5% Distilling Claude Opus 4.6, GPT-5.5, Gemini 3.1 Pro Into Qwen3

huggingface.co · 2026-07-27 · Score : 18

UT Austin's ISO Paper Frames RLVR as 'Inherit the Spectrum, Optimize the Frames,' Hits AdamW Accuracy in 2.7x Fewer Steps

huggingface.co · 2026-07-21 · Score : 23

Loopie MoE looped transformer claims IMO and IPhO gold-medal performance without tools

huggingface.co · 2026-07-19 · Score : 18

Muon optimizer lifts agentic RL success from 0.29 to 0.55 on ALFWorld, new arXiv paper finds

huggingface.co · 2026-07-20 · Score : 21

LoRA Speedrun Launches Public Wall-Clock Leaderboard for Fine-Tuning Techniques on Qwen2.5-1.5B and SmolLM2-1.7B

github.com · 2026-07-19 · Score : 17

Mind Lab Ships LongStraw — GRPO-Based Long-Context RL Training Reaches 4.46M Token Positions on 8 H20 GPUs

huggingface.co · 2026-07-16 · Score : 25

Tsinghua's SEED Frames Agentic RL as Self-Evolving On-Policy Distillation, Adds 14.9-45.9 Points on ALFWorld Over GRPO

huggingface.co · 2026-07-16 · Score : 19

Ring-Zero Scales Zero-RL to a Trillion Parameters, Shows Emergent Multi-Step Reasoning Without Handcrafted Signals

arxiv.org · 2026-07-16 · Score : 27

HF Paper 'Proxy Exploration and Reusable Guidance': Modular LLM Post-Training Paradigm Uses Proxy-Guided Update Signals, Separates Exploration From Policy Training for Cheaper Sample-Efficient RL Fine-Tuning

huggingface.co · 2026-07-13 · Score : 15

Tsinghua's SAO Paper: Single-Rollout Async Optimization Cracks GRPO's Stability Wall — Trained the Open GLM-5.2 750B-A40B, Beats Baselines on SWE-Bench Verified and BeyondAIME

huggingface.co · 2026-07-08 · Score : 18

BottleCap AI Releases ThinkingCap-Qwen3.6-27B — Reasoning Fine-Tune That Cuts Thinking Tokens 50% on Average (74% on GSM8K) While Preserving Accuracy and Safety

huggingface.co · 2026-07-06 · Score : 17

Waterloo Paper 'Program-as-Weights' Compiles Natural-Language Specs Into Compact Adapters — 0.6B Qwen3 Interpreter Matches Direct-Prompted Qwen3-32B at ~1/50th Inference Memory, Ships FuzzyBench 10M-Example Dataset

arxiv.org · 2026-07-03 · Score : 16

Paper 'Is One Layer Enough?' Shows Training a Single Transformer Layer Can Match Full-Parameter RL Training — Gains Concentrate in a Small Middle-Layer Subset

arxiv.org · 2026-07-01 · Score : 19

HF Paper 'TRIAGE': Role-Typed Credit Assignment Adds Semantic Role Axis to GRPO — Boosts ALFWorld Success 7.9% on Qwen2.5-7B, Cuts Rollout Length 10–15%

huggingface.co · 2026-06-30 · Score : 17

HF Paper 'QVal' — Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents Gains Traction

huggingface.co · 2026-06-30 · Score : 16

OPID: On-Policy Skill Distillation for Agentic RL Gains +9 Points on ALFWorld and +11 Points on WebShop Over GRPO Without External Skill Libraries or Privileged Inference Context

huggingface.co · 2026-06-25 · Score : 18

ShutterMuse: Fudan University and StepFun Train Unified MLLM on 130K-Sample Dataset for Real-Time Photography Composition and Pose Guidance — Introduces CaptureGuide-Bench

huggingface.co · 2026-06-25 · Score : 17

Adaption Launches AutoScientist — AI Tool That Co-Optimizes Training Data and Model Weights in a Single Loop, Claims Win-Rate Doublings Across Models

techcrunch.com · 2026-05-13 · Score : 24

Unsloth and NVIDIA Publish Three LLM Training Optimizations Delivering ~25% Additional Speedup at Zero Accuracy Cost

unsloth.ai · 2026-05-06 · Score : 21

Numéros de la newsletter (16)

#517 What Happens When AI Runs Out of Content to Steal?

2026-07-29 · 1 mentions

World Models: $610M Bets the World Is Not a Token

2026-06-21 · 1 mentions

AI Safety, Security & Ethics: train one trait, get alignment free

2026-06-21 · 3 mentions

AI Regulation: export-control law just became an AI kill switch

2026-06-21 · 3 mentions

Deep Learning: 4-bit pretraining just matched FP8 at scale

2026-05-26 · 1 mentions

AI Regulation: Trump killed his own AI oversight order at the door

2026-05-26 · 2 mentions

Frontier Research: LLMs reach Nature as the field bets against them

2026-05-26 · 1 mentions

AI Safety: the deepfake takedown clock starts ticking

2026-05-26 · 4 mentions

Frontier Research: reasoning is leaving the token stream

2026-05-21 · 6 mentions

AI Careers: 11,000 jobs cut in a day, one role is booming

2026-05-21 · 1 mentions

AI Safety: Washington floats an FDA for AI models

2026-05-21 · 3 mentions

Frontier research: the next architecture is not transformer-shaped

2026-04-25 · 1 mentions

ML this week: open weights closed the gap from underneath

2026-04-25 · 3 mentions

Deep Learning: Mamba-3 halves the state. SSMs win again.

2026-04-18 · 1 mentions

Generative AI: Anthropic just retook the frontier, quietly

2026-04-18 · 1 mentions

Machine Learning: The dataset just became the alignment lever

2026-04-18 · 1 mentions