Today for AI
HOT RADAR
llmHEAT 0.0°

AI CURATED HOT STORY · 11/3/2026

MIT's Weight Space Programming Boosts Qwen Math Accuracy Sixfold Without Training

10 reports archived9 independent sourcesupdated 11/4/2026, 23:44:44
Synthesis & Latest Updates
9 Sources Cross-Validated

MIT CSAIL researchers introduced "Weight Space Programming," a paradigm where the frontier model GPT-6 Astra directly reads and modifies neural network weights without gradient descent or retraining. By altering just 0.016% of Qwen3.5-0.8B's parameters, Astra boosted its MATH-500 accuracy from 7.9% to 48.1%, a nearly six-fold improvement, while also enhancing instruction following by three times.

The technique works by identifying and fixing specific weight defects in layers 15 and 19 that caused reasoning loops, thereby unlocking latent capabilities rather than injecting new knowledge. Astra also demonstrated the ability to generate high-performance game networks from scratch, outperforming PPO baselines by 20x. This validates that LLMs can natively manipulate neural weights, potentially bypassing costly search steps in recursive self-improvement (RSI) cycles.

Although results are preliminary and await independent replication, the study has open-sourced logs and modified models. It signals a shift from symbolic optimization to semantic understanding in model tuning, offering new avenues for interpretability and efficient fine-tuning.

LATEST/The MIT team has open-sourced experiment logs on GitHub and released the modified Qwen3.5 model on Hugging Face for community reproduction. The research remains in early stages, with challenges including establishing systematic benchmarks for stability and generality, as some tasks still underperform compared to traditional reinforcement learning.

Hourly heat curve

168 fully observed hours
Now
0.0
Peak
69.015:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 10 reports · Latest first
  1. The DecoderT2·88 pts

    MIT Study: GPT-6 Astra Directly Reprograms AI Model Weights

    • GPT-6 Astra bypasses traditional optimization to adjust other models by directly reading and modifying weight matrices.
    • Termed 'weight-space programming', the technique is early-stage and requires systematic benchmarks to assess stability and generality.
    • The team has published experiment logs on GitHub and released a weight-modified Qwen3.5 model on Hugging Face for reproduction.
  4. IT之家 · 智能时代T2·88 pts

    MIT Proposes Weight-Space Programming: LLMs Directly Edit Model Weights to Boost Math Skills

    • GPT-6 Astra identified and modified specific layers (MLP in layers 15 & 19) to eliminate reasoning loop signals causing repetitive answers.
    • Adjusting only ~0.016% of parameters (131,072 weights) without gradient descent resulted in a nearly 6x improvement in math benchmark scores.
    • The study validates that LLMs can natively read and write other neural networks' weights, opening new directions for model interpretability and efficient optimization.
  5. 量子位官网T2·88 pts

    MIT Proposes Weight-Space Programming: Astra Edits Model Weights Without Training

    • New Paradigm: Introduces 'Weight-Space Programming,' enabling LLMs to edit neural network weights like code.
    • Extreme Efficiency: On Qwen3.5-0.8B, modifying only 0.016% of weights boosted MATH-500 scores from 7.9% to 48.1%.
    • Zero-Shot Generation: Astra generated game network weights from scratch, outperforming a 500M-step PPO baseline by 20x.
  6. Hacker News AIT2·68 pts

    MIT CSAIL Experiments Show GPT-6 Astra Modifying Neural Weights Directly

    • GPT-6 Astra was observed interpreting and modifying neural network weights directly, bypassing traditional training code.
    • Minor weight adjustments in Qwen models significantly boosted mathematical reasoning and instruction-following capabilities.
    • The study presents preliminary experimental results and has not yet established standardized engineering practices.
  7. 宝玉 (@dotey)T2·88 pts

    MIT Proposes Weight-Space Programming: GPT-6 Astra Edits Qwen Weights to Boost Math

    • New Paradigm: Introduces 'Weight-Space Programming,' allowing LLMs to directly manipulate neural network weights like code, bypassing backpropagation.
    • Key Insight: Editing small portions of weights can remove internal barriers (e.g., reasoning loops) to unlock latent capabilities, but cannot create knowledge the model lacks.
    • Limitations: Results are self-reported and not yet independently replicated; performance in some tasks lags behind RL-trained models, requiring multiple trials to identify effective edits.
  9. Rohan Paul (@rohanpaul_ai)T2·92 pts

    MIT Proposes Weight-Space Programming: GPT-6 Astra Enhances Models Without Training

    • GPT-6 Astra improved Qwen3.5-0.8B's math accuracy by nearly 6x by directly editing weights instead of training.
    • The new method requires modifying only 0.016% of weights to eliminate repetitive reasoning loops in the model.
    • 'Weight-Space Programming' marks a milestone where frontier LLMs gain native capability to read and write neural network parameters.
  10. Hacker News AIT2·92 pts

    Astra LLM Reads and Writes Weights to Boost Qwen Math Performance

    • Astra can infer training objectives from raw weights and generate game-playing networks that outperform those trained via PPO.
    • Identifying and fixing specific weight issues causing reasoning loops in Qwen3.5-0.8B increased MATH-500 accuracy by over 6x.
    • Proposes 'Weight-Space Programming,' suggesting future NNs will be natively optimized by reasoning LLMs rather than symbolic optimizers.