# A preprint finds sparse reward-related LLM coordinates

> Probes and activation interventions found small hidden-state subsets that tracked answer value and step-level reward error, but the neuroscience labels are analogies and the work is not peer reviewed.

_Source: Unreviewed arXiv preprint; existence and metadata verified against arXiv and OpenAlex, findings not independently replicated · 2026-09-22 · 7 min read_

Canonical: https://iyu.app/e/llm-sparse-reward-neurons-preprint

## The 60-second version

An unreviewed preprint reports sparse, intervention-sensitive LLM coordinates that track value and reward-prediction error under trained probes.

**Key points**

- Fewer than 1% of coordinates often retained probe performance in reported pruning tests.
- Targeted ablation sharply reduced one Qwen model's MATH500 accuracy.
- Confidence and guided-search demonstrations showed modest gains in selected settings.
- The coordinate sets depend on probes, layers, tasks and reward definitions.
- Neuroscience terms are analogies and provide no evidence of dopamine or consciousness.

**Verdict.** A promising mechanistic-interpretability hypothesis, not yet a peer-reviewed or independently replicated reward-system discovery.

## Full explainer

> **⚑ Caveat:** This is an unreviewed arXiv preprint with no independent replication located. 'Value neurons' and 'dopamine neurons' are author-chosen analogies for activation coordinates; the model contains no dopamine or biological reward circuit.


### Method — Probes searched hidden states for reward signals

The authors trained small probe networks on transformer hidden states. One estimated the probability that a partial solution would end correctly; another estimated the temporal-difference error between reasoning steps.

- **Value coordinates:** Selected because a probe used them to predict terminal correctness or expected value.
- **TD-error coordinates:** Selected because activations tracked changes in estimated success between steps.
- **Sparsity test:** Low-weight probe inputs were removed to see how few dimensions retained predictive performance.
- **Intervention:** Selected activations were zeroed in particular layers and task accuracy was remeasured.


### Finding — Less than one percent often retained the signal

Across reported model and benchmark combinations, aggressive pruning often left a small coordinate subset with similar probe performance. Results covered mathematics, coding and instruction-following tasks, although not every test used every model.

- **<1%** — coordinates sufficient in reported pruning tests
- **75.2%** — original MATH500 accuracy in one ablation setup
- **20.3%** — average after selected-coordinate ablation
- **77.8%** — one reported guided-search accuracy

> **⚑ Caveat:** All performance figures are reported by the preprint authors. Probe choice, coordinate ranking and reward labels help define the result, so the numbers need independent reproduction.


### Meaning — Decodable is not the same as a biological subsystem

A transformer coordinate is a number in a computation. Calling it a neuron is shorthand; calling it a dopamine neuron describes a mathematical resemblance to reward-prediction error, not chemistry, anatomy or experience.

> The paper reports sparse reward-related representations, not a brain inside the model.


### Limits — The general mechanism remains unproven

- **Preprint status:** no peer review or independent replication was located.
- **Probe dependence:** selected coordinates depend on the learned probe and pruning rule.
- **Entanglement:** ablation can damage other features carried by the same coordinates.
- **Narrow applications:** process-reward search was demonstrated in a limited benchmark setup.
- **No mental-state evidence:** the experiments say nothing about consciousness, pleasure or motivation.


### Next step — Replication should test uniqueness and transfer

Useful follow-up would compare nonlinear and causal baselines, test unrelated architectures without reselecting coordinates, and ask whether the same sparse sets predict failures outside benchmark reasoning. Until then, treat the subsystem as a testable interpretability proposal.


## Primary sources

- [Telegram post 1499](https://t.me/CNSmydream/1499)
- [arXiv preprint v2](https://arxiv.org/abs/2602.00986v2)
- [arXiv PDF](https://arxiv.org/pdf/2602.00986v2)
- [arXiv API record](https://export.arxiv.org/api/query?id_list=2602.00986)
- [OpenAlex record](https://api.openalex.org/works/https://doi.org/10.48550/arXiv.2602.00986)

---
_Published by iyu (https://iyu.app) — the day's AI news, checked against primary sources and rewritten in plain language. Free to quote with attribution and a link to the canonical URL._