Weight-space programming claims draw independent replications and critiques
Frontier models reportedly editing small networks' weights without gradients sparks debate, while Tencent’s Hy4 leads Agent Arena and new open-weight releases emerge.
1 source · cross-referenced
- Frontier LLMs may directly generate or edit small networks' weights without gradient training, with multiple independent replications reported.
- Critics argue some claims lack competent baselines or overfit to benchmarks, while supporters frame the work as meta-learning progress.
- Tencent’s Hy4 leads Agent Arena at $0.55 per task, and open-weight releases include Kimi K3.1 Pro, P-H3-Image, and LongCat-2.5.
- Artificial Analysis updates capability and intelligence indexes, with Claude Opus 5.5 leading in most categories.
A series of experiments claim frontier large language models can generate or edit weights of small neural networks directly, without gradient-based training or external tools. Multiple independent groups have replicated aspects of these findings, including a reproduction by Hugging Face using its ML Intern research harness and a user-reported hand-written classifier outperforming a trained version.
One reconstruction showed that GPT-6 Astra edited 131,072 weights in Qwen3.5-0.8B without backpropagation, raising MATH-500 scores from 7.9% to 48.1%. However, a narrower Zhihu reconstruction attributed much of the gain to fixing a 32K-token loop failure mode rather than recursive self-improvement. A laptop reproduction by Simon Willison confirmed the weight-interpretation feat on Qwen 3.8 27B after 45+ minutes of runtime.
Critics argue the headline claims lack competent baselines and that some results reflect benchmark overfitting, such as fitting weights directly to a benchmark without meaningful generalization. Others question whether models are primarily reading or manipulating only the first and last layers, which map to known inputs and outputs. Supporters counter that the work demonstrates meta-learning finally working and suggest hybrid approaches that interleave backpropagation with direct weight edits.
In agentic evaluations, Tencent’s Hy4 model reportedly scores 75.2 on APEX-Agents versus 72.1/72.9 for DeepSeek V4.1 Flash-1017, though the two trade wins on other benchmarks. Hy4 ranks 14th overall across more than 8,500 Agent Arena sessions at $0.55 per task, compared to $1.46 for Kimi K3.1 Pro (Max).
New open-weight releases include Kimi K3.1 Pro, with weights scheduled for release on Wednesday; Pruna’s P-H3-Image, a distilled model running in 8 steps instead of 50 and producing a 4 MP image in 2.6 seconds; and LongCat-2.5, a 1.6T total parameter model with approximately 48B active parameters priced at $0.30/$1.20 input/output tokens.
Artificial Analysis updated its Capability Indexes v2.0 to emphasize non-hallucination and agentic tool use, with Claude Opus 5.5 (max) leading all six indexes. The Intelligence Index v5.0.1 showed no frontier ranking change greater than one point after regrading with the CritPt v2 server. In robotics, Opus 5.5 completed a robot-arm task and then spent 22 additional minutes on it, while only GPT-6.1 Sol and Gemini 4 Argon stopped within a minute.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74