Benchmark: can an LLM make a language model better at math by editing its weights directly? (Qwen3.5-0.8B, MATH-500/GSM8K, via OpenRouter)
-
Updated
Nov 3, 2026 - Python
Benchmark: can an LLM make a language model better at math by editing its weights directly? (Qwen3.5-0.8B, MATH-500/GSM8K, via OpenRouter)
Agent skill for weight-space programming: edit LLM weights
Open-source weight space editing for small language models. Direct tensor edits, measured hill climbs and a Hugging Face workbench.
SGNE-CC: training-free 3.13 MiB norm-preserving weight edit for Qwen3.5-0.8B vs. MIT-p50c across 5 benchmarks at 32k tokens (38 JSONL traces & CPU verifier)
Give AI agents tools to inspect, edit, and test model weights: an MCP harness and Python toolkit for weight-editing experiments.
Model engineering: AI agents edit a model's internals in small, signed, reversible diffs. Research programme.
Weight-space programming on a 27B production model: sandboxed kernel, metered forward passes, sealed held-out eval, every transcript
Claude agents repair a thinking loop in Qwen3.5-0.8B by editing its weights directly (rank-1 edits, no training). Pre-paper, code, lab notebook and data.
Make your own small AI that runs on a laptop: teach it your documents or describe how it should answer. Honest report cards, fixes by editing weights, one-file export for Ollama.
Direct weight surgery for LLMs — edit weights, not gradients. Extract the arrow, leave the body intact.
Signet: fixing a language model by editing its weights with plus-or-minus steps only. Qwen2.5-0.5B, full sealed-test record.
Closed-form weight edits for Gemma 4 that a coding agent can drive: find where the model makes a choice, build low-rank edits without gradients, and judge each against size-matched random edits (target, keep, side effects). MLX and PyTorch backends; exports LoRA adapters.
Fixing small language model defects (reasoning loops…) by having an LLM (Claude Opus in Claude Code) edit their weights directly, without training — measured against the original by slm-quality-bench
Can frontier models fix a coding agent by editing its weights? Code, protocols, records, frozen gate reports and agent configs of the Gemma 4 E4B weight-edit study, with the full paper (PDF) and both Kaggle writeups.
Can an LLM repair sycophancy by editing weights directly? A gradient-free edit of 0.23 % of Qwen2.5-0.5B cuts wrong-hint flips from 76.8 % to 32.3 % on a blind test split, but correct hints also help less on unknown questions.
The map of modifying model weights directly: knowledge editing, steering & abliteration, task arithmetic, unlearning, and the weight-space programming wave. 权重编辑全景 · 中英双语 · Daily radar 🤖
WeightX: Open-source Weight-Space Programming (WSP) harness in Rust & Tauri 2 + PyTorch sidecar for safe, declarative and verified LLM weight surgery.
To associate your repository with the weight-editing topic, visit your repo's landing page and select "manage topics."