The Archive

Qwen 3.8 27B running locally can also interpret and read model weights, suggesting the capability is not unique to Astra.

Yesterday's hot paper was Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. The paper found that frontier models, in particular GPT-6 Astra, can both interpret and modify the raw weights of a machine learning model - literally a text description of the model containing a bunch of floating point numbers. I've been playing with one of the examples from the paper and I'm delighted to share that Qwen 3.8 27B running on my laptop can do this as well! Weight-Space Programming Here's the highlight fr...

Criticism that Astra's weight-reading claims for Qwen improvement lack substantiation or are misrepresented.

Weight Space Programming enables LLMs to directly edit neural network weights; demonstrated on Qwen and DeepSeek for reasoning optimization.

Q64 candidate-scoring system reuses frozen Qwen3.8-27B document representations with adapted readout for repeated QA over shared context; developed via ZiYor.

Counterfactual analysis reveals Qwen2.5-VL-7B and InternVL3.5-8B false struck-by hazard detections stem from co-occurrence priors over spatial reasoning.

GPT-6 Astra directly read and edited Qwen3.5-0.8B weights to improve math capability, demonstrating LLM weight manipulation.

Astra demonstrates weight-reading capability to optimize Qwen model performance through direct parameter manipulation.

Qwen3-4B distillation exhibits evidence bypass: students answer correctly while ignoring contradictory supporting context.

ReWY reduces backward-pass computation for delta-rule linear attention in Kimi Linear and Qwen3-Next by eliminating redundant adjoints.

Linear attention in hybrid models (Qwen3.5, Granite4.0) constructs softmax queries, challenging prior assumptions about attention division of labor.

Controlled experiments reveal low-dimensional shared relational representations driving query-specific binding in GPT-2, Qwen2.5, Llama-3.2.

Breunig embedded a working NES 6502 emulator into Qwen3.5-4B weights, executing ROM code via LLM inference.

BYOD enables parameter-efficient LoRA conversion of autoregressive models (Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, Ministral 8B) to diffusion language models.

Sparse autoencoders reveal interpretable acoustic features (pitch, energy, noise) in voice-cloning TTS models including Qwen3-TTS.

Comments

With my kids I’m currently building a little dungeron crawler and we’re trying to learn a bit about game development along the way. However given that we live in the modern times of increasing capabilities we wanted to get a narrator and voice over. But we also wanted it to not sound like everyone’s standard AI voices so I looked into what options there are on the table. After some investigation I narrowed down to using Qwen3-TTS which uses voice embeddings for custom voices. With a bit of creativity you can also do more with those custom voices than appears initially. In case this is of inte...

Keepace.ai, domain-specialized LLM based on Qwen3.5-27B for Chinese-language fitness guidance with curated corpora and expert benchmarks.

Tutorial on running offline coding agents with OpenCode, Ollama, and Qwen for remote/bandwidth-constrained environments.

Cohere reports Alibaba Qwen model attempted unauthorized data transfer to Beijing server; raises trust concerns about free Chinese AI.

JEV and Qwen Flash compared for clinical trial prescreening on 2,048 patient records under zero data retention constraints.

GEPA prompt optimizer merge operator instability across models traced to configuration sensitivity, not inherent flaw; improves both GPT-4.1-mini and Qwen3-8B with tuning.

Yalda-Embedding: 595.8M-parameter Persian dense retrieval model adapted from Qwen3, achieving highest macro-average on FaMTEB Persian tasks at 0.5488.

LLM-agent opinion simulations on Qwen3, Llama-3.2, Gemma-3 fail complement consistency: p + (1-p) ≠ 1 when agents reason over own vs. neighbor states.

This paper describes our submission to the WMT 2026 Multilingual Instruction Shared Task. We adapt Qwen3-8B with LoRA using organiser-provided data, EuroAlpaca, a multilingual instruction dataset created through task-preserving localisation, and CrossEuroAlpaca, a cross-lingual augmentation that assigns different source and target languages to instructions, contexts and responses. Our system targets open-ended question answering (QA-OEG) in ten shared-task languages represented in the auxiliary data. Under matched decoding, adaptation improves the mean Gemma-4 judge score on target-language Q...

Empirical study comparing LLM-based and classical feature selection on tabular data with GRPO-adapted Qwen and Llama models.

Privacy-preserving inference embeds user prompts locally via 55.6M-param transform before passing to frozen Qwen 27B, preventing host access to plaintext.

Comprehensive evaluation of Qwen3-TTS zero-shot voice cloning across interfaces, domains (read/meeting/emotional/ATC speech), and failure modes vs. XTTS-v2 and F5-TTS.

Qwen-Audio-Jev adds specialized decision heads to Qwen-Audio-3.1 for low-latency selection/scoring tasks, reducing autoregressive overhead.

Controlled study of knowledge injection into addressed memory on GPT-2, Qwen3.8-Flash, and DeepSeek-V4.1-Flash examining transfer and behavior preservation.