LLMs can read neural network weights. But do they actually understand what they're reading?
I tested this on 9B–30B language models.
The results surprised me.
A paper from @Poindexter_Labs, inspired by Weight-Space Programming:
Introducing Weight-Space Reading👇
[AI] Latent Feedback Is a Compute and Data Amplifier: Scaling Law for Latent Feedback Transformers on Repeated Data
B Huang, C Dwivedi, X Ma… [IMF & CMU] (2026)
https://t.co/1txZ3YwNA2
Sometime ago I had proposed an antidote to the hype in robotics demo tapes released on X - a live challenge where a child competes with a robot to finish a task, and we measure performance by speed of completion, weighted by success. I would like to congratulate @roeiherzig and collaborators for making this happen at CoRL 2026. Two robot policies competed and as expected they were trounced by the child. But this is fine! Progress will come and one day robots will exceed humans across the board. In computer vision the famous PASCAL object detection challenge started in 2005 with performance around 5%, and look where we are now.
Training with fewer bits is faster, but quantization error can interfere with learning.
Today on Connectionism, we detail how we trained Inkling 2 Small with a native 4-bit forward pass from the first step of pretraining through the final RL rollout.
https://t.co/CIMECSEpD0
We are typically interested in distilling models.
Our new work distills datasets.
Dataset Distillation via Drifting is a new method that can "summarize" an entire dataset into a handful of images.
The images look slightly unnatural, but this is by design: the entangle concepts from multiple real samples into one synthetic sample.
Models trained on these images perform well in downstream tasks, such as classification.
A fun artsy application: you can use our method to "summarize" an artist's work into one painting.
[Read our paper and George's thread for more details]
Want to try weight-space programming? Today we're releasing dspy-wsp, a prototype weight-space optimizer that works like any other @DSPyOSS optimizer.
Using dspy-wsp, and about ~$5-10 of Astra, we improved every decision model we tested. https://t.co/iPsOTkNZym
This one weird trick improves alignment generalization! Just add random noise during training, telling the model to keep their values stable under perturbation
Reprogramming dormant neurons : Qwen TTS
Ever wondered why Qwen3-TTS sometimes breathes or laughs out of nowhere? Agent found neurons behind it.
The details:
Agent found 417 neurons to be specific (about 0.48% of the talker's MLP neurons) and 61 of them are dormant. Amplified together, they give the following effects effect:
1. Breathes as words are spoken
2. Hushed ASMR
3. Natural laughs as language is being spoken
4. Enhanced whispering
5. 2nd speaker appearing.
Comparisons and neuron map repo in the comments below.
Inspired by recent work on Lazarus by @DimitrisPapail , which reprogrammed dormant neurons.
Last couple years I've been asked what I think about decoder-only vs encoder-decoder for multimodality, and never really given an answer.
Ellis and Chris take a good step towards answering that question in this study.
Two big points:
(1) Recent encoder-free models like Gemma4-12B, Inkling1, TUNA-2, etc are all a single "point estimates": they show this is possible but tell us absolutely nothing about how it scales, or doesn't! -> this paper does.
(2) Encoder-free or not is actually a false dichotomy. It hides three questions which Chris and Ellis nicely disentangle, and answer separately:
(a) separate pre-training,
(b) separate vision params, and
(c) vision params before or inside decoder?
Answer in pic below, and full thread and paper.
Lean Verified Precision (w/ @Robertljg)
Recently, "A Fix for Softmax" demonstrated a surprising bug in the textbook softmax implementation. How do we avoid having subtle bugs in the world's most expensive software? 🧵
🍌 What if an image generator could see?
@GoogleDeepMind's Vision Banana showed it can, but kept data, recipe and weights closed. So we rebuilt it fully open:
✨ FLUX.2 Klein 4B/9B finetune w/ public data only
✨ 73.9 vs Vision Banana’s 69.9 mIoU zero-shot on Cityscapes
✨We release code + 148 checkpoints for the community/
Aaand we innovate:
📈 We develop a new test-time scaling recipe we call “legend-ensembling”: this gives strong gains by applying more compute without retraining anything.
🔗 https://t.co/lwI4PYcBLn
📄 https://t.co/Jhyxqco54u
A @FunAILab work from A. Jadon, @v_pariza, @FragileGoodwill, @jonaruthardt, @ASophiaKoepke, @abursuc and me.
I gave a guest lecture at CMU on speech quality assessment. It's an update of our INTERSPEECH 2025 tutorial.
Slides are now publicly available here:
https://t.co/vvFgJxWbtx
Thank you for the invitation, Shinji! @shinjiw_at_cmu
we figured out RL for speech recognition
At @OrukLabs we stopped giving wrong words positive credit.
Our full recipe cut mean WER from 4.936% to 4.757% across nine public English test sets.
Open weights below.
PLR: Plackett-Luce for Reordering In-Context Learning Examples
The order of few-shot examples changes LLM accuracy, but n! orderings cannot be searched. PLR fits a Plackett-Luce distribution over orderings that give high accuracy.
https://t.co/dH1jAPjpps