Qwen3.5-0.8B weights were edited by GPT-6 Astra, changing only 0.016% of parameters and fixing reasoning loops. The edit boosted math performance up to 6× and instruction following up to 3× without gradients or training.
Key points
Weight change: 0.016% of Qwen3.5-0.8B weights altered.
Performance gains: math reasoning improved up to 6×; instruction following improved up to 3×.
Paper available on Papers with Code.
Sources
- Original post
- paperswithcode.co - Linked in the post
GPT-6 Astra edits Qwen3.5-0.8B weights, improves math 6x
GPT-6 Astra edited Qwen3.5-0.8B model weights, changing only 0.016% of them and boosting math performance up to 6× while increasing instruction following 3×, all without gradients or training.
Key points
Weight change: 0.016% of Qwen3.5-0.8B weights were altered.
Performance gains: math reasoning improved up to 6×; instruction following improved 3×.
Sources
3D World Simulator runs at 50+ FPS on RTX 5090
3D World Simulator is an interactive world model that runs at 50+ FPS on an RTX 5090. It predicts touch as 3D particles aligned with scene geometry and generates novel view appearance, multi-view consistent RGB-D, 3D geometry, and spatially aligned tactile data.
Key points
Runs at 50+ FPS on an RTX 5090.
Predicts touch as 3D particles alongside scene geometry.
Sources
Hydrogen-1.0 launch with benchmark scores
Hydrogen-1.0 is introduced as an open-weight model for scientific research, scoring 38.3% on FrontierScience-Research and 93.0% on GPQA-Diamonds. The model was post-trained GLM-5.3 on internal scientific data, lab records, and discovery trajectories.
Key points
Benchmark: 38.3% on FrontierScience-Research vs GPT-6-Astra’s 33.3%
Benchmark: 93.0% on GPQA-Diamonds
Training data: internal scientific data, lab records, and discovery trajectories
Model base: post-trained GLM-5.3
Sources
Astra weight edit improves Qwen3.5-0.8B math 6x
Astra edited Qwen3.5-0.8B weights, changing 0.016% of parameters, fixing reasoning loops and boosting math performance up to 6× and instruction following up to 3×, without gradients or training. The technique is called Weight-Space Programming.
Key points
Weight edit changed only 0.016% of Qwen3.5-0.8B weights.
Math performance improved up to 6×; instruction following improved up to 3×.
Sources
GPT-6 Astra edits Qwen3.5-0.8B weights, improves math 6×
GPT-6 Astra edited Qwen3.5-0.8B model weights, changing 0.016% of them. The edit fixed reasoning loops and boosted math performance up to six times and instruction following up to three times, without using gradients or training.
Key points
Weight change: 0.016% of Qwen3.5-0.8B parameters were edited.
Performance gains: math reasoning improved up to 6×; instruction following improved up to 3×.
Sources
Qwen3.5-0.8B weight edit improves math 6×, instruction 3×
Qwen3.5-0.8B weights were edited by GPT-6 Astra, changing 0.016% of parameters and yielding up to 6× better math performance and 3× better instruction following, all without gradients or training. The method is called Weight-Space Programming.
Key points
Changed only 0.016% of weights.
Math performance improved up to 6×; instruction following improved 3×.
Sources
ChefNMR2 – new approach for structure elucidation from NMR
ChefNMR2 is a new approach for structure elucidation from NMR spectra. The work is led by @zyxiong123 and a paper is linked online.
Key points
ChefNMR2: new approach for structure elucidation from NMR spectra
Paper: chefnmr.cs.princeton.edu
Sources
- Original post
- chefnmr.cs.princeton.edu - Linked in the post
