Shanghai AI Lab authors report big benchmark gains from editing a few model weights
A paper posted on November 9, 2026 studies “weight-space programming” on Qwen3.5-0.8B and Qwen3.5-9B. The authors report that editing only 1,023 parameters raised MATH-500 accuracy by 29.45 points.
What changed
The authors tested 16 editing trajectories across seven benchmarks. They report gains of up to 29.45, 4.42, 36.72 and 17.11 points on MATH-500, GPQA Diamond, IFEval and SciBench.
Why it matters
If the results hold, small direct edits to weights could be a cheap way to adjust a model without full retraining.
What remains uncertain
These are the authors’ own results on small models and are not independently reproduced. Six of the 16 trajectories kept the original weights because gains could not be reliably validated.
What to watch
Watch for replication on larger models and other benchmarks.
Source
Why this ranks here
An interesting method with large reported gains, but small models and no outside check yet.
- Impact5
- Reach5
- Novelty7
- Commitment4
- Coverage quality5
- Future potential6