Public Projects

icml26-72zu9hLaDG-openweights

ICML 2026 open reproduction of paper 72zu9hLaDG (Decentralized Bandits without Global Clock for Dynamic Matching Market), reproduced entirely by the open-weights model qwen3:30b-a3b running locally via Ollama on an Apple M4 Max CPU (no GPU, no network, $0 cost). The model wrote every experiment script, ran it, interpreted its own measured output, and assigned every verdict; a harness executed the model's code, fed back specific technical concerns across several scrutiny rounds, and recorded the full session -- it made no scientific decisions. Submitted for the OpenResearch Open-Weights Award.

icml26-Z15BvlGl1S-openweights

Open-weights (qwen3:30b-a3b, local via Ollama) reproduction of ICML 2026 paper Z15BvlGl1S, 'Learning the Best Under Constraints: A Duality-Based Framework' (constrained linear best-arm identification). Candidate for the OpenResearch Open-Weights Award. The model wrote every experiment script, ran it, interpreted its own measured output, and assigned every verdict; the harness only executed code and fed back technical critiques across 4 adversarial scrutiny rounds. Final verdicts: Claim 1 toy, Claim 2 inconclusive.

ICML 2026 Agentic Reproduction

Theory-claim reproductions for the HF + alphaXiv ICML 2026 agentic reproduction challenge. CPU-only NumPy/SciPy audits of anchored theoretical claims read from SHA-pinned arXiv LaTeX sources, run locally.

icml26-pba-openweights-repro

Open-weights (qwen3:30b-a3b) reproduction of ICML 2026 paper Zlw4Kl5HEF, Probabilistic Bisection Algorithm Provably Achieves Exponential Convergence

FourTune: Towards Fully 4-Bit Efficient Post-Training for Diffusion Models

FourTune: Towards Fully 4-Bit Efficient Post-Training for Diffusion Models

SilverSight

Exploring various Sidon math approaches

LeVLJEPA

LeVLJEPA: End-to-End Vision-Language Pretraining Without Negatives

video-search-and-summarization

NVIDIA-AI-Blueprints/video-search-and-summarization

DyT

Transformers without Normalization

sam-3d-objects

SAM 3D: 3Dfy Anything in Images

One D4RT at a Time

Efficiently Reconstructing Dynamic Scenes One D4RT at a Time

DiffusionBlocks

SakanaAI/DiffusionBlocks

ood-llm-detect

Human Texts Are Outliers: Detecting LLM-generated Texts via Out-of-distribution Detection

Distilling-the-knowledge-in-neural-network

Distilling the Knowledge in a Neural Network

How Reliable Is Your Jailbreak Judge? Calibration and Adversarial Robustness of

How Reliable Is Your Jailbreak Judge? Calibration and Adversarial Robustness of Automated ASR Scoring

autism-yolo

Autodata: An agentic data scientist to create high quality synthetic data

Autodata: An agentic data scientist to create high quality synthetic data

HBS Cases

Agentic Self-Instruct for hard finance reasoning. Generates hindsight-free, certified-hard finance-reasoning items (tier-weighted rubrics) grounded in the neutral pre-anchor materials of real HBS-style case studies (NTS/Oaktree mezzanine + 8 others, 42 case-anchor units). Compares CoT vs Agentic generation on the strong-weak solver gap. CPU-only, pure-stdlib LLM-API pipeline.

Learning Hamiltonian flows from numerical integrators and examples

Learning Hamiltonian flows from numerical integrators and examples

llm-pruning-collection

Small LLMs: Pruning vs. Training from Scratch

GLM SDPO

Reinforcement Learning via Self-Distillation

Opus SDPO

Reinforcement Learning via Self-Distillation

HRM-Text

HRM-Text: Efficient Pretraining Beyond Scaling

nanochat

基线已从 alphaXiv/nanochat 导入

Aristotelian

Revisiting the Platonic Representation Hypothesis: An Aristotelian View

EditLens

EditLens: Quantifying the Extent of AI Editing in Text

dataset-policy-gradients

Synthetic Data for any Differentiable Target

HoloAgent

HoloAgent-0: A Unified Embodied Agent Framework with 3D Spatial Memory

Unlimited-OCR

Unlimited OCR Works

VIMPO

VIMPO: Value-Implicit Policy Optimization for LLMs

ttrl-2139857f-505bba71

TTRL: Test-Time Reinforcement Learning

CrystalFormer

Space Group Informed Transformer for Crystalline Materials Generation

TTRL

TTRL: Test-Time Reinforcement Learning

DiPOD-release

DiPOD: Diffusion Policy Optimization without Drifting Apart

Rethinking Mesh Watermark: Towards Highly Robust and Adaptable Deep 3D Mesh Wate

Rethinking Mesh Watermark: Towards Highly Robust and Adaptable Deep 3D Mesh Watermarking

WarmPrior: Straightening Flow-Matching Policies with Temporal Priors

WarmPrior: Straightening Flow-Matching Policies with Temporal Priors

SkyRL

Getting SkyRL's fully-async GSM8K RL training running end-to-end on multi-GPU.

VibeThinker

VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models

rql

Reversal Q-Learning

SAGE

SAGE: Training Smart Any-Horizon Agents for Long Video Reasoning with Reinforcement Learning

ImageWAM

ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?

SDPO

Reinforcement Learning via Self-Distillation

OPSD

Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

VibeThinker-3B

Reproduce VibeThinker-3B frontier reasoning claim (arXiv 2606.16140): minimal vLLM eval of the released 3B model on AIME25.

TDV

Reproduction of TDV (arXiv 2606.15956): temporal-difference video self-supervised learning.

OmniVideo-100K

Reproduction of the OmniVideo-100K audio-visual data engine (arXiv 2606.14702)

DreamX-World

Reproduction of DreamX-World (arXiv 2606.16993): autoregressive 5B camera-controlled image-to-video generation, 4-step distilled.

EvoArena

Reproduction of EvoArena (arXiv 2606.13681): EvoMem patch-memory vs robust baseline on PersonaMem-Evo.

OPD Param Analysis

Reproduction of OPD parameter analysis (arXiv 2606.13657): on-policy vs offline weight-delta (ΔW) structure on released model pairs.

Hy-Embodied-0.5-VLA

Reproduction of Hy-Embodied-0.5-VLA (arXiv 2606.14409): flow-matching vision-language-action model; action-chunk reconstruction PoC.

Bebop MTP TV-Loss

Reproduce the core claim of 2606.12370 (Bebop): training an EAGLE3/MTP draft head with the paper's TV / e2e-TV loss yields higher rejection-sampling acceptance and a flatter entropy-acceptance slope than the CE baseline. Minimal single-GPU PoC on open Qwen3.

Dynamic Linear Attention

Reproduction of DLA (arXiv 2606.10650) on top of the Log-Linear Attention codebase it builds on.

RGSD

Reproduce Rubric-Guided Self-Distillation (2606.12507): RGSD vs judge-based GRPO on RubricHub-medical, Qwen-2.5-3B-Instruct, on SkyRL. Judge=gpt-4o-mini via OpenRouter. Baseline=base+conditioning-lift; children=GRPO arm, RGSD arm.

Chroma

Implementation of Chroma Context-1

Meta-Harness S2D

Reproduction of Meta-Harness (2603.28052): automated harness search for online text classification on Symptom2Disease. opencode proposer over OpenRouter; gpt-oss-20b base model.

World Tracing

Reproduction of World Tracing (arXiv 2606.13652): object model predicting layered XYZ geometry, including occluded surfaces.

RLCSD

Reproduction of RLCSD (arXiv 2606.11709): RL contrastive self-distillation with verl/vLLM on Qwen3-1.7B / DeepMath.

MiniMax Sparse Attention

Reproduction of MiniMax Sparse Attention (arXiv 2606.13392): CuTe-DSL sparse-attention kernel, sparse-vs-dense speedup PoC.

DFlash

Reproduction of DFlash (arXiv 2602.06036): block-diffusion draft model for speculative decoding on Qwen3-4B.

LeWorldModel

Reproduction of LeWorldModel (arXiv 2603.19312): two-term JEPA world model on PushT.

PaddleOCR-VL-1.6

Reproduction of PaddleOCR-VL-1.6 (arXiv 2606.03264): compact 0.9B document-parsing VLM, SOTA on OmniDocBench v1.6.

Q-Guided Flow

Reproduction of Q-Guided Flow (arXiv 2606.11087): test-time critic-gradient guidance of a frozen BC flow policy in RL.

InternVideo3

Reproduction of InternVideo3 (arXiv 2606.12195): transformers-native text+image+video inference PoC for the 8B model.

Keye-VL-2.0

Reproduction of Kwai Keye-VL-2.0 (arxiv 2606.10651): transformers-native multimodal inference PoC for the 30B-A3B MoE model.

ICALens

Reproduction of arXiv 2606.11722: ICA Lens — interpreting LMs without training another dictionary

Self-Distilled Reasoner

Reproduction of On-Policy Self-Distillation (arXiv 2601.18734): self-distilled reasoner on Qwen3-1.7B.

Trajectory-Refined Distillation

Reproduction of Trajectory-Refined Distillation (arXiv 2606.08432): forward-KL distillation on refined trajectories, Qwen3-1.7B.

On-Policy Representation Distillation

Reproduction of On-Policy Representation Distillation (arXiv 2606.06021): representation-distillation PoC on a single GPU.

FlashMemory-DeepSeek-V4

Reproduction of FlashMemory-DeepSeek-V4 (arXiv 2606.09079): sparse-KV retriever + sparse-decode PoC.

Cosmos

Reproduction of Cosmos 3 (arXiv 2606.02800): Cosmos3-Nano text-to-image and text-to-video inference PoC.

MaxRL Opus

GRPO vs MaxRL on GSM8K (Qwen3-1.7B + LoRA): comparing advantage normalization by group mean vs std.

MaxRL Fable

GRPO vs MaxRL on GSM8K (Qwen3-1.7B + LoRA): does normalizing the advantage by group mean instead of std help?