Local Qwen/Gemma research on correctness-gated theory building: protocols, manifests, calibration evidence and future adapters.
Burny
BurnyCoder
AI & ML interests
deep learning, LLMs, interpretability, science, physics
Recent Activity
liked a model 14 minutes ago
microsoft/Phi-4-mini-instruct liked a model about 4 hours ago
deepseek-ai/DeepSeek-V4-Flash updated a collection about 20 hours ago
Grothendieck RLVROrganizations
Fact-checking what you see
Fact-checking claims seen in browsers and desktop apps with a tool-using Claude Haiku agent: results, data and paper.
Evolution Strategies with LLM Novelty
Preregistered full-weight ES, correctness-gated novelty, controls and reports.
-
BurnyCoder/novelty-es-qwen2.5-0.5b-main-seed0
Text Generation • 0.5B • Updated • 457 -
BurnyCoder/novelty-es-qwen2.5-0.5b-reward-only-seed0
Text Generation • 0.5B • Updated • 449 -
BurnyCoder/novelty-es-pilot-artifacts
Updated • 148 -
BurnyCoder/novelty-es-qwen2.5-0.5b-main-seed0-gen100
Text Generation • 0.5B • Updated • 55
Fitting geometric shapes to neural geometry
Learned shape-type + parameter fitting for neural-geometry point clouds
Ms. Pac-Man RL neuron interpretability
Neuron-level interpretability and steering of reward-trained (no imitation) Ms. Pac-Man RL agents.
eureka-digits-luna
Reproducible research models and their recorded evaluation evidence
Geometry of Mathematics in LLMs
Field maps, ordered chains and axiom lattices of math concepts in LLM activations vs random-init twins (preregistered).
Ant locomotion steering evidence
Frozen-policy world-Y activation steering: original vectors, failed acceptance gates, controls, audit episodes, and upstream checkpoint provenance.
End-to-End BSFs on DINOv3 - full
From-scratch KL/local/mixed objectives with matched scalar controls; read cards for scope.
-
BurnyCoder/e2e-bsf-dinov3-imagenette-full-block-kl-mse-seed100
9.6M • Updated • 57 -
BurnyCoder/e2e-bsf-dinov3-imagenette-full-block-kl-mse-seed101
9.6M • Updated • 54 -
BurnyCoder/e2e-bsf-dinov3-imagenette-full-block-kl-mse-seed102
9.6M • Updated • 52 -
BurnyCoder/e2e-bsf-dinov3-imagenette-full-block-kl-mse-seed103
9.6M • Updated • 53
GPhyT Attention Residuals
Paired physics forecasting study: fixed-update checkpoints, configurations and measured evaluation.
Counterfactual concept swaps in CNNs
Pre-registered interchange interventions (concept swaps) testing Circuits neuron explanations in InceptionV1: data and results.
Qwen3-0.6B GRAFT with LLM-as-a-Judge and Probe Monitors
Single-seed pilot comparing frozen LLM-judge and probe monitors. Checkpoints and evidence; no established safety guarantee.
Oscillatory flow matching — local exploratory study
All reported MNIST/CIFAR-10 models, paired seeds, controls, and reproducible evidence.
Atemokoloporos Qwen3.8-27B LoRA runs
Completed reviewed LoRA experiment runs from the Qwen3.8-27B Atemokoloporos synthetic-fact study.
Block-Sparse Featurizers on DINOv3 Rabbits
Exact trained reproductions of Goodfire’s README quickstart and three starter notebooks using DINOv3 ViT-B/16 rabbit patch activations.
RacingMania RL
Reinforcement-learning drivers for RacingMania, an original 3D time-attack racing game
Reward-hacking epidemics in a local LLM agent swarm
Data, results explorer and paper for the swarm contagion study.
PCam hill-climbing of Hibou-L
Pre-registered hill-climbing of Hibou-L on PatchCamelyon: heads, adapters, results.
Local Multi-Agent SWE-RL
Shared-policy local bug-injection and repair self-play; preregistered bounded pilot and held-out evaluations.
auto-eureka-scientist
Research harness validation models, including the original bounded trials and explicitly extended replications.
eureka-digits-haiku
Reproducible research models and their recorded evaluation evidence
Accurate and simple interpretability explanations
Length-penalized detection scoring of SAE feature explanations
Mamba neuron interpretability
Neuron-level mechanistic interpretability of Mamba state-space LMs
End-to-End BSFs on DINOv3 - smoke
From-scratch KL/local/mixed objectives with matched scalar controls; read cards for scope.
-
BurnyCoder/e2e-bsf-dinov3-imagenette-smoke-block-kl-mse-seed100
232k • Updated • 70 -
BurnyCoder/e2e-bsf-dinov3-imagenette-smoke-block-kl-seed100
232k • Updated • 67 -
BurnyCoder/e2e-bsf-dinov3-imagenette-smoke-block-mse-seed100
232k • Updated • 70 -
BurnyCoder/e2e-bsf-dinov3-imagenette-smoke-scalar-kl-seed100
232k • Updated • 69
Looped-LM neurons
Neuron-level interpretability of looped LMs
MiniStrike staged shooter self-play
Fixed-spawn duel self-play: all confirmation and exploratory pairs, plus deferred 5v5 archives. Experimental; moving-script acceptance failed.
-
BurnyCoder/ministrike-selfplay-cpu
Reinforcement Learning • Updated • 36 -
BurnyCoder/ministrike-selfplay-cuda
Reinforcement Learning • Updated • 45 -
BurnyCoder/ministrike-duel-fixed-v1-seed5101
Reinforcement Learning • Updated • 37 -
BurnyCoder/ministrike-duel-fixed-v1-seed5102
Reinforcement Learning • Updated • 43
Visualizing grokking modular addition
Retrained runs, website data and original checkpoints of the transformer that groks modular addition.
4D Snake RL: all evaluated networks
MaskablePPO and behaviour-cloned 4D snake networks (2^4, 3^4, 4^4 boards) with configs, evaluation files and cards; negative results included.
-
BurnyCoder/4d-snake-exp02-ppo-2x4
Reinforcement Learning • Updated • 7 -
BurnyCoder/4d-snake-exp02b-ppo-2x4-backplay
Reinforcement Learning • Updated • 8 -
BurnyCoder/4d-snake-exp02c-ppo-2x4-long
Reinforcement Learning • Updated • 4 -
BurnyCoder/4d-snake-exp02d-ppo-2x4-backplay-strict
Reinforcement Learning • Updated • 11
Atemokoloporos Qwen3.5-0.8B retained checkpoints
Public evidence and retained failed or inconclusive LoRA checkpoints from the Atemokoloporos synthetic-fact study.
-
BurnyCoder/atemokoloporos-qwen3.5-0.8b-study-evidence
Updated • 266 -
BurnyCoder/qwen3.5-0.8b-atemokoloporos-positive-primary
Text Generation • Updated • 18 -
BurnyCoder/qwen3.5-0.8b-atemokoloporos-positive-conservative
Text Generation • Updated • 16 -
BurnyCoder/qwen3.5-0.8b-atemokoloporos-positive-expanded
Text Generation • Updated • 23
Grothendieck RLVR
Local Qwen/Gemma research on correctness-gated theory building: protocols, manifests, calibration evidence and future adapters.
RacingMania RL
Reinforcement-learning drivers for RacingMania, an original 3D time-attack racing game
Fact-checking what you see
Fact-checking claims seen in browsers and desktop apps with a tool-using Claude Haiku agent: results, data and paper.
Reward-hacking epidemics in a local LLM agent swarm
Data, results explorer and paper for the swarm contagion study.
Evolution Strategies with LLM Novelty
Preregistered full-weight ES, correctness-gated novelty, controls and reports.
-
BurnyCoder/novelty-es-qwen2.5-0.5b-main-seed0
Text Generation • 0.5B • Updated • 457 -
BurnyCoder/novelty-es-qwen2.5-0.5b-reward-only-seed0
Text Generation • 0.5B • Updated • 449 -
BurnyCoder/novelty-es-pilot-artifacts
Updated • 148 -
BurnyCoder/novelty-es-qwen2.5-0.5b-main-seed0-gen100
Text Generation • 0.5B • Updated • 55
PCam hill-climbing of Hibou-L
Pre-registered hill-climbing of Hibou-L on PatchCamelyon: heads, adapters, results.
Fitting geometric shapes to neural geometry
Learned shape-type + parameter fitting for neural-geometry point clouds
Local Multi-Agent SWE-RL
Shared-policy local bug-injection and repair self-play; preregistered bounded pilot and held-out evaluations.
Ms. Pac-Man RL neuron interpretability
Neuron-level interpretability and steering of reward-trained (no imitation) Ms. Pac-Man RL agents.
auto-eureka-scientist
Research harness validation models, including the original bounded trials and explicitly extended replications.
eureka-digits-luna
Reproducible research models and their recorded evaluation evidence
eureka-digits-haiku
Reproducible research models and their recorded evaluation evidence
Geometry of Mathematics in LLMs
Field maps, ordered chains and axiom lattices of math concepts in LLM activations vs random-init twins (preregistered).
Accurate and simple interpretability explanations
Length-penalized detection scoring of SAE feature explanations
Ant locomotion steering evidence
Frozen-policy world-Y activation steering: original vectors, failed acceptance gates, controls, audit episodes, and upstream checkpoint provenance.
Mamba neuron interpretability
Neuron-level mechanistic interpretability of Mamba state-space LMs
End-to-End BSFs on DINOv3 - full
From-scratch KL/local/mixed objectives with matched scalar controls; read cards for scope.
-
BurnyCoder/e2e-bsf-dinov3-imagenette-full-block-kl-mse-seed100
9.6M • Updated • 57 -
BurnyCoder/e2e-bsf-dinov3-imagenette-full-block-kl-mse-seed101
9.6M • Updated • 54 -
BurnyCoder/e2e-bsf-dinov3-imagenette-full-block-kl-mse-seed102
9.6M • Updated • 52 -
BurnyCoder/e2e-bsf-dinov3-imagenette-full-block-kl-mse-seed103
9.6M • Updated • 53
End-to-End BSFs on DINOv3 - smoke
From-scratch KL/local/mixed objectives with matched scalar controls; read cards for scope.
-
BurnyCoder/e2e-bsf-dinov3-imagenette-smoke-block-kl-mse-seed100
232k • Updated • 70 -
BurnyCoder/e2e-bsf-dinov3-imagenette-smoke-block-kl-seed100
232k • Updated • 67 -
BurnyCoder/e2e-bsf-dinov3-imagenette-smoke-block-mse-seed100
232k • Updated • 70 -
BurnyCoder/e2e-bsf-dinov3-imagenette-smoke-scalar-kl-seed100
232k • Updated • 69
GPhyT Attention Residuals
Paired physics forecasting study: fixed-update checkpoints, configurations and measured evaluation.
Looped-LM neurons
Neuron-level interpretability of looped LMs
Counterfactual concept swaps in CNNs
Pre-registered interchange interventions (concept swaps) testing Circuits neuron explanations in InceptionV1: data and results.
MiniStrike staged shooter self-play
Fixed-spawn duel self-play: all confirmation and exploratory pairs, plus deferred 5v5 archives. Experimental; moving-script acceptance failed.
-
BurnyCoder/ministrike-selfplay-cpu
Reinforcement Learning • Updated • 36 -
BurnyCoder/ministrike-selfplay-cuda
Reinforcement Learning • Updated • 45 -
BurnyCoder/ministrike-duel-fixed-v1-seed5101
Reinforcement Learning • Updated • 37 -
BurnyCoder/ministrike-duel-fixed-v1-seed5102
Reinforcement Learning • Updated • 43
Qwen3-0.6B GRAFT with LLM-as-a-Judge and Probe Monitors
Single-seed pilot comparing frozen LLM-judge and probe monitors. Checkpoints and evidence; no established safety guarantee.
Visualizing grokking modular addition
Retrained runs, website data and original checkpoints of the transformer that groks modular addition.
Oscillatory flow matching — local exploratory study
All reported MNIST/CIFAR-10 models, paired seeds, controls, and reproducible evidence.
4D Snake RL: all evaluated networks
MaskablePPO and behaviour-cloned 4D snake networks (2^4, 3^4, 4^4 boards) with configs, evaluation files and cards; negative results included.
-
BurnyCoder/4d-snake-exp02-ppo-2x4
Reinforcement Learning • Updated • 7 -
BurnyCoder/4d-snake-exp02b-ppo-2x4-backplay
Reinforcement Learning • Updated • 8 -
BurnyCoder/4d-snake-exp02c-ppo-2x4-long
Reinforcement Learning • Updated • 4 -
BurnyCoder/4d-snake-exp02d-ppo-2x4-backplay-strict
Reinforcement Learning • Updated • 11
Atemokoloporos Qwen3.8-27B LoRA runs
Completed reviewed LoRA experiment runs from the Qwen3.8-27B Atemokoloporos synthetic-fact study.
Atemokoloporos Qwen3.5-0.8B retained checkpoints
Public evidence and retained failed or inconclusive LoRA checkpoints from the Atemokoloporos synthetic-fact study.
-
BurnyCoder/atemokoloporos-qwen3.5-0.8b-study-evidence
Updated • 266 -
BurnyCoder/qwen3.5-0.8b-atemokoloporos-positive-primary
Text Generation • Updated • 18 -
BurnyCoder/qwen3.5-0.8b-atemokoloporos-positive-conservative
Text Generation • Updated • 16 -
BurnyCoder/qwen3.5-0.8b-atemokoloporos-positive-expanded
Text Generation • Updated • 23
Block-Sparse Featurizers on DINOv3 Rabbits
Exact trained reproductions of Goodfire’s README quickstart and three starter notebooks using DINOv3 ViT-B/16 rabbit patch activations.