Summary
The public scrape was empty again, an eleventh straight day with no reachable Nitter instance, so zero curated reposts and zero posts from the tracked AI handles. The saved-reading path was not empty, and that is the whole of this morning's signal: one new bookmark landed overnight, and it is the third save in two days after a stretch where the feed authenticated cleanly and returned nothing new for over a week. The save is a screenshot of the first page of Sparse Reward Subsystem in Large Language Models, from Tsinghua and Stanford, which finds that the reward-related information already known to live in a model's hidden states is concentrated in a sparse set of individual neurons splitting into two functional types, and that one of those types works as a step-level scorer for inference-time search. That is a cost result wearing interpretability clothes, and it is the strongest thing in this window. It is also four months old and reached the reader through a post that got the attribution, the headline number and the recency all wrong, which is worth recording because the same account produced the other overnight save with the same distortion pattern. There is no cluster here in the usual sense, one save is one save, but read with yesterday evening's two it forms a coherent three-item run: attention cost, reward-signal cost, and adoption dynamics. Everything else this morning came through Gmail and the research feed rather than social.
Posts
Sparse reward neurons inside LLMs, saved as a screenshot of the paper's first page (@HowToPrompt__, paper: arXiv 2602.00986). The attached image is the title block and abstract of Sparse Reward Subsystem in Large Language Models by Guowei Xu of Tsinghua with Mert Yuksekgonul and James Zou of Stanford, legible enough to read in full. The actual claim: prior work established that a model's hidden states encode reward-related information such as answer correctness and confidence, but it read that signal with black-box probes fitted on the entire hidden-state vector, which shows the information is decodable and says nothing about where it lives. This paper does simple probing at the neuron level and finds the signal concentrated in a sparse subset, splitting into two types. Value neurons predict state value, the expected probability that continuing generation from the current point yields a correct answer. Dopamine neurons encode step-level temporal-difference error, the surprise when that expectation jumps or drops. The names are borrowed from neuroscience because biological value and dopamine neurons encode the same two quantities. The paper reports value neurons as robust and transferable across datasets and models with causal evidence, and then cashes it out in two applications: value neurons as a confidence predictor, and dopamine neurons as a process reward model guiding inference-time search. That last one is why this matters for cost. A process reward model, the scorer that grades each intermediate reasoning step rather than only the final answer, is normally a separate trained model you serve alongside the policy, so every candidate step in best-of-N or tree search costs two forward passes instead of one. If the signal is already sitting in a handful of the policy's own activations, the second model becomes a gather over known indices. The tweet's framing is wrong in three specific ways and each is worth naming: it says "Stanford discovered," dropping the Tsinghua first author; it claims ablating a small fraction of value neurons collapses math reasoning by over 50 percent, a number that does not appear in the abstract; and it presents this as new, when the arXiv identifier is 2602 and HuggingFace surfaced it on 2026-05-11. It also adds a biological-inevitability story, "nobody programmed this," to a paper that explicitly predicts the finding in advance from a maximum-entropy reinforcement learning argument, since an optimal policy and its soft Q-function are coupled and a strong policy therefore has to be encoding value information to make good next-token decisions at all. → Wiki summary
Nothing from the public timeline. No curated reposts and no posts from the tracked AI handles in this window. The general scrape has produced nothing since 08-27.