cd /news/large-language-models/localization-does-not-imply-steerabi… · home › topics › large-language-models › article
[ARTICLE · art-165167] src=machinebrief.com ↗ pub= topic=large-language-models verified=true sentiment=· neutral

Localization Does Not Imply Steerability: A Mechanistic Study of Sign Errors in LLM Mathematics

A preregistered mechanistic study of six open models from 14B to 675B parameters found that wrong sign tokens in multi-step LLM mathematics are committed late, by MLP layers in the last quarter of the network (70-100% per model and task), and that subtracting a single sign direction from the residual stream flips 13 of 14 held-out minus-written integration-by-parts errors on Llama-3.3-70B, with flips replicating on Qwen, Gemma and Mistral-Small at 24-70% per cell. The authors report 23 of 29 preregistered checks passed, with pass criteria frozen in git before held-out data was seen, and all misses reported. The study concludes that localization does not imply steerability: Phi-4 implements an equally well-localized circuit at the same late depth, but the identical intervention flips almost nothing because the sign is written redundantly across roughly 15 layers with margins several times larger (median 17.8 logits).

by read1 min views1 publishedNov 6, 2026

arXiv:2611.07478v1 Announce Type: cross Abstract: When large language models do multi-step mathematics, a characteristic failure recurs: the model writes one wrong sign token mid-computation -- a cofactor sign, the minus of a recursive integration-by-parts step -- then executes the rest of the arithmetic correctly, and the answer comes out wrong. We ask where inside the network that sign is written. Across six open models spanning 14B to 675B parameters, a per-layer readout shows the wrong sign is committed late -- by MLP layers in the last quarter of the network (70-100% per model and task), a pattern present, where tested (Qwen), in the base model before instruction tuning. The mechanism varies: some models carry fixed "minus-stamp" layers that push a minus regardless of the right answer, others carry amplifiers of whichever sign was already chosen, and Qwen adds a downstream rescue layer that opposes the stamp. The circuit is causal: subtracting a single sign direction from the residual stream flips 13 of 14 held-out minus-written integration-by-parts errors on Llama-3.3-70B, with flips replicating on Qwen, Gemma, and Mistral-Small (24-70% per cell), and with representation-level specificity -- the same dose breaks correctly-minus answers while sparing correct-plus answers where the direction is surgical (0/72 on Llama), and matched random-direction and control-layer interventions are null. Core claims were preregistered with pass criteria frozen in git before held-out data was seen: 23 of 29 checks passed, all misses reported. But causality is not uniform: Phi-4 implements an equally well-localized circuit at the same late depth, yet the identical intervention flips almost nothing -- the sign is written redundantly across ~15 layers with margins several times larger (median 17.8 logits). Localization, we conclude, does not imply steerability: two models can share a circuit location yet differ enormously in how overturnable it is.

── more in #large-language-models 4 stories · sorted by recency
── more on @llama-3.3-70b 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/localization-does-no…] indexed:0 read:1min 2026-11-06 · —