Astra can read and write weights to improve Qwen

Who Should Read

An LLM app developer who wants to fix loops or weaknesses in a small open model without fine-tuning.

Core Mechanics

• LLM named the game from weights alone
  • GPT-6 Astra identified the trained game from 17,253 weights.
  • Game options were not given; Breakout was 99% confident.
• It wrote weights from scratch with no training
  • It beat PPO trained for 100M steps in all three MinAtar games.
  • Breakout score was 20x PPO at 500M steps.
• Fixing 0.016% removed Qwen's loops
  • Astra edited 0.016% of Qwen3.5-0.8B's weights to fix reasoning loops.
  • With the same sampling settings, MATH-500 rose from 7.9% to 48.1%.
• The math fix carried over to instruction following
  • The same edit raised IFEval by 3x.
  • No training loop was used.
• It fixed a stuck policy in one shot
  • On the Kinetix car task it edited 3.5% of the weights.
  • Success rose from 0% to 100% over 1,024 episodes.
• Larger networks got tools and forward passes
  • The 1.5M-parameter Go network used a Python sandbox.
  • With 50 forward passes it beat an equal-sized PPO net trained on 20M games.

Community Reactions

• Reproduction attempt showed no gain
  • Someone edited 1,920 weights of a natural-language-to-bash small model with Astra.
  • Some grep mistakes were fixed but other commands broke.
• Critique that two results are conflated
  • A commenter said the small-model results were read-only with no tools.
  • The Qwen experiment used tools and looked like standard steering vector techniques.
• Qwen 2B experience cut token waste
  • A commenter reported about 20% fewer tokens wasted before reasoning starts.
  • They had not yet reproduced a loop after the fix.
• View that the claims are overstated
  • A commenter said game guessing relied on shallow weight patterns.
  • The language task was a 3-option choice, so significance may be overstated.

How to Apply

• Ask Astra to diagnose your model today
  • Write your small model's symptom, such as loops, in one sentence.
  • Paste the model card warnings and a few failure samples and ask for likely causes.
• Compare before and after on a fixed eval set
  • Freeze at least 50 failure cases before touching any weights.
  • Also check whether other commands or features broke after the edit.
• Try it only when symptoms concentrate in one area
  • Apply it to problems like loops whose cause narrows to a single direction.
  • Do not use it as a broad performance booster.

Try It

# Paste into a strong LLM chat such as Astra or Claude (simplified version of the paper's method: diagnosis only, no weight editing)

I use the small model {{model name}}. Symptom: {{e.g., repeats the same sentence and never finishes an answer}}

Failure samples:
{{3-5 failing outputs}}

Question: Split the possible causes into 3 and design an experiment for each that needs no weight edits. Then give a hypothesis on which part of the weights might be fixed.

Next step: run one suggested experiment and paste the results back to narrow down the hypothesis.

Terminology

Related Papers

• Show HN: Qwen3-4B with the AI writing tics trained out

  Distilling Qwen3-4B with LoRA cut its AI-writing slop score from 121 to 12.5 per 1000 words

• Dust: Pretraining Transformers Without Backpropagation

  Adding per-token noise to activations lets transformers pretrain without backpropagation, rivaling backprop.

• Show HN: TurboGPT: train 22KiB transformer in 13s

  A CUDA C++ project trains a tiny byte-level GPT in under a minute.

• I had Gemini train its own replacement for $9

  Fine-tuning GLiNER on $9 of Gemini labels reached 0.83 F1 locally.

• Learning to solve hard problems in RL for LLMs by never giving up

  Cutting compute on easy problems and retrying hard ones in RL raised accuracy on the hardest problems

• Training a 3.8B LLM to 0.384 CORE for $998

  A 3.8B LLM trained on 8 B200s in 43 hours for $998 scored 0.384 on CORE.

• I were 17, I'd learn how to build LLMs from scratch

  PG says at 17 he'd skip startups and learn to build LLMs from scratch.

• Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship. Try it

  Distilling DeepSeek V4 Flash into GPT-OSS on finance boosted skill but didn't transfer censorship

Related Resources

• https://rikiyat.github.io/blog/weight-space-programming/
• https://arxiv.org/abs/2110.13100
• https://arxiv.org/abs/2505.12470
• https://arxiv.org/abs/2402.18153