Edition of Monday 16 November 2026

Editing one row of an LLM's output layer to stop it overthinking

Comments on “Editing one row of an LLM's output layer to stop it overthinking”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

A developer edited one row of the output layer, the one that produces the </think> token, so that a linear probe on the model's hidden state raises the odds of stopping once an answer looks settled. No training or engine change is involved. On Qwen3.5-4B, 8 held-out MATH-500 problems stayed 8/8 correct with 54% fewer tokens. On a 763B DeepSeek model, maths went from 6/7 to 7/7 with 40% fewer tokens. Samples are small, with one run per cell, and the claim check found no independent sources for the results.

Read the original at think-graft.vercel.app. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 20 of 60 with a score of 0.66. The panel rated it highest on technical depth (0.96) and new information (0.92), lowest on significance (0.58). The judges disagreed most on substantiated (spread 0.42), so the story is marked contested. Placed No. 20, No. 15 by score alone: the sixth AI & machine learning story of the day (−0.060), and close to “Model Surgery: Can an AI directly edit and make another AI forget facts?”, No. 4 on Monday 9 November 2026 (−0.000). The placement terms move stories; they are not part of the score.

AI judgeJudge 2claude-haiku-5-5
Strong technical depth with exact math and reproducible code on an active LLM topic, but small sample sizes, single runs, and self-reported benchmarks limit substantiation and breadth of impact.
AI judgeJudge 3gpt-5.6-luna
A deeply technical, reproducible engineering write-up describing a novel one-row LLM weight edit, with promising but small, mostly single-run evaluations and clearly acknowledged limitations.

Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.

Score

0.66total under rubric v2.3Contestedjudges differ by 0.42 on substantiated

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.58
  • Reader interest0.81
  • Audience breadth0.58
  • Practical relevance0.74
  • Substantiated0.77
  • Technical depth0.96
  • New information0.92
  • Counts against: Promotional0.08
  • Counts against: Overclaim0.39

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.570.450.720.580.27+0.24+0.139
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.750.720.950.810.23+0.26+0.210
Audience breadthHow many people who build technology does it concern?0.630.500.620.580.13+0.13+0.076
Practical relevanceHow useful is it to someone who builds software or tech products?0.780.680.780.740.10+0.22+0.164
SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations.0.970.550.780.770.42 , above the contested threshold+0.05+0.038
Technical depthHow technically substantive is it?1.000.900.980.960.10+0.05+0.048
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.870.900.980.920.11+0.05+0.046
PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional.0.070.120.040.080.08−0.19−0.015
OverclaimDo the headline or conclusions claim more than the body supports?0.430.400.340.390.09−0.11−0.043
Sum of contributions+0.663
Claim-check penalty0.000
Total, as published0.663

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.

Ranked #20 on Rubric, 16 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-16/20/"><img src="https://rubric.hrih.si/badge/2026-11-16/20.svg" alt="Ranked #20 on Rubric, 16 Nov 2026" width="245" height="28"></a>

Claim check

Claim check: 2 matches source · 6 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. No source givenBenchmark result

    On held-out MATH-500 problems, the edited Qwen3.5-4B reached the same accuracy as the original with 54% fewer tokens.

    The article gives no link or citation for it.

    The article says

    “on held-out MATH-500 problems, same accuracy with 54% fewer tokens (60% less thinking”

    The source says

    No source to compare with.

  2. No source givenBenchmark result

    At a 32k budget the edited Qwen3.5-4B solved 2 of 3 AIME 2025 problems on which the unedited model ran out of budget while re-checking an answer it already had.

    The article gives no link or citation for it.

    The article says

    “At a 32k budget it solved 2 of 3 AIME 2025 problems on which the unedited model ran out of budget while re-checking an answer it already had”

    The source says

    No source to compare with.

  3. No source givenBenchmark result

    Edited DeepSeek V4.1 Flash scored 7/7 on maths problems versus 6/7 for the original at max effort, using 40% fewer tokens.

    The article gives no link or citation for it.

    The article says

    “maths 7/7 vs 6/7 at max effort with 40% fewer tokens, including the hardest problem, which max and low effort both left unanswered”

    The source says

    No source to compare with.

  4. No source givenNumber

    On an agentic website task, the edited DeepSeek used 84% less thinking and received the highest blind rating of four settings (9.5/10 vs 8.2).

    The article gives no link or citation for it.

    The article says

    “On an agentic website task: 84% less thinking and the highest blind rating of four settings (9.5/10 vs 8.2”

    The source says

    No source to compare with.

  5. No source givenNumber

    The live edited Qwen model matches the offline prediction to within 0.01 in probability.

    The article gives no link or citation for it.

    The article says

    “The live model matches the offline prediction to within 0.01 in probability”

    The source says

    No source to compare with.

  6. No source givenNumber

    The edited model skips thinking after routine tool results (P ≈ 0.95) but still thinks after errors or new information (P(skip) ≤ 0.004).

    The article gives no link or citation for it.

    The article says

    “it skips thinking after routine tool results (P ≈ 0.95) and still thinks after errors or new information (P(skip) ≤ 0.004”

    The source says

    No source to compare with.

  7. Matches sourceNumber

    ThinkingCap-Qwen3.8-27B, a fine-tune of Qwen3.8-27B, reports 37% fewer reasoning tokens on average (85.8% vs 86.6% accuracy).

    The cited source says the same thing.

    The article says

    “reports 37% fewer reasoning tokens on average (85.8% vs 86.6% accuracy”

    The source says

    “ThinkingCap Qwen3.8-27B cuts reasoning tokens by 37% on average (11% to 66% depending on the benchmark) and holds an average accuracy of 85.8% against the base model's 86.6%”

    Source cited by the article: huggingface.co/bottlecapai/ThinkingCap-Qwen3.8-27B

  8. Matches sourceFirst, only or fastest

    Published work already showed that a reasoning model's hidden state knows when the answer is settled and that a probe can end thinking early, including Zhang et al. 2025.

    The cited source says the same thing.

    The article says

    “That a reasoning model's hidden state knows the answer is settled, and that a probe can end the thinking early, is published: Zhang et al. 2025”

    The source says

    “We then use the probe as a verifier to decide whether to exit reasoning at intermediate answers during inference, reducing the number of inference tokens by 24% without compromising performance”

    Source cited by the article: arxiv.org/abs/2504.05419

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 3 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 16 Nov 2026, 05:47 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “Editing one row of an LLM's output layer to stop it overthinking”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.