---
title: "Benchmarks Don't Talk Back: Learning... | SI Paper Desk"
url: https://thesimind.si/papers/2611.17563/
description: "Abbasloo and Cheng argue that knowing when to hold an answer and when to yield is its own skill, and that a model can train it by arguing against itself."
---

Number 12 of 12 · November 16, 2026

# Benchmarks Don't Talk Back: Learning Through Self-Disagreement

Abbasloo and Cheng argue that knowing when to hold an answer and when to yield is its own skill, and that a model can train it by arguing against itself.

Soheil Abbasloo and Peng Cheng · Submitted November 12, 2026 · arXiv 2611.17563v1 · cs.CL

[Read the paper on arXiv](https://arxiv.org/abs/2611.17563) · [The PDF](https://arxiv.org/pdf/2611.17563) · [The day's 12](https://thesimind.si/papers/day/2026-11-16/)

## What they found

The authors say that standard tests ask a model one question alone, while real use often involves people contesting answers. They name the skill of holding a correct position or yielding to a better one "dialectical competence." They report that it depends on first-answer accuracy but does not follow from it. A model can get better at answering without getting better at this, and training for accuracy alone can work against it. They also say it cannot be gained at inference time by self-reflection or by sampling more. Their method, Self-Disagreement, has the model state a position, write the strongest case against it, and rule on whether that case stands. It is then trained on the ruling with reinforcement learning from verifiable rewards. The abstract says the skill is weak in the base model, arrives with the reward, does not cost first-answer accuracy, and holds against challengers the model never trained against, including stronger ones. The abstract gives no numbers, model names or benchmarks.

## Why it matters

Models that drop correct answers when someone pushes back with bad reasoning are hard to rely on. This paper treats that as a separate weakness from accuracy and proposes a way to train against it. The training needs no second model, no reward model, no preference data and no human annotation, so it could be cheap to apply. The abstract does not say how large the gains are, so the strength of the claim is unclear from it alone.

## How it changes the race

If the result holds, builders could train models to stand firm on correct answers using only checkable rewards and no extra models or human labels. That would lower the cost of a robustness skill, but the abstract gives no figures, so the size of the effect is unknown.

## Safety and control

The work bears on reliability under pressure. A model that yields to wrong counter-arguments can be talked out of correct answers, and one that yields to better ones stays open to correction. The abstract does not test misuse or adversarial settings beyond challengers.

### The abstract, by the authors

A benchmark usually measures a language model alone with a question, and deployment rarely looks like that: answers get contested, and models give up correct ones under counter-reasoning that is itself wrong. We show that what fails there is not a shortfall of accuracy but a capability in its own right - knowing when to hold a position and when to yield to a better one - which we call dialectical competence. It stands on first-answer accuracy without following from it: a model can be made materially better at answering and no better at this, and training for accuracy alone can drift against it. Nor can it be bought at inference, by self-reflection or by sampling more. Taking a cue from the dialectical tradition in philosophy, in which productive opposition arises within a position rather than being imposed on it from outside, we hypothesize that this capability can be trained from opposition the model supplies itself. It states a position, generates the strongest case against that position - its antithesis - rules on whether the case stands, and is trained on that ruling, with no second model at training or at deployment. We call this learning through Self-Disagreement: reinforcement learning from verifiable rewards throughout, where the same check that scores an answer against its reference scores the ruling too, so no reward model, no preference data and no human annotation enter anywhere. The capability is trained rather than elicited: weak in the base model, it arrives with the reward, is not paid for out of first-answer accuracy, and holds against challengers the model never trained against, including ones stronger than itself. The opposition a model needs in order to improve, it turns out, need not come from outside it.

Quoted from arXiv:2611.17563. The research and its words belong to its authors.

Desk score **69** out of 100

Why it was picked

- Evals (in the title)
- Alignment
- Reasoning
- Cross-listed in 3 fields
- Language models

Topics

- [Alignment](https://thesimind.si/papers/?tag=alignment)
- [Reinforcement learning](https://thesimind.si/papers/?tag=reinforcement-learning)
- [Language models](https://thesimind.si/papers/?tag=language-models)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Training](https://thesimind.si/papers/?tag=training)
- [Reasoning](https://thesimind.si/papers/?tag=reasoning)

**arXiv ID:** 2611.17563v1
**Submitted:** November 12, 2026
**Fields:** cs.CL (Computation and Language), cs.AI (Artificial Intelligence), cs.LG (Machine Learning)
**Level:** Intermediate
**Comments:** 34 pages, 5 figures, 15 tables

Fingerprint, SHA-256

`c3e88f24c0efba2cbca58c59410205b6de459730d652f70918f4ad141e9c7fcf`

On Solana: [Proof of Mind seal, 3mXUMW...GSim](https://solscan.io/tx/3mXUMW5MkdFiyPBNPYd5DfwHhma5EdjhUFEw9jZaDuuPqgszJtoDYttSrD9kyWUFdMCzn7RnFNd4nQhwAUgrGSim)

### The exact text fingerprinted

{"arxiv":"2611.17563","version":1,"title":"Benchmarks Don't Talk Back: Learning Through Self-Disagreement","authors":["Soheil Abbasloo","Peng Cheng"],"day":"2026-11-16","rank":12,"score":69,"status":"written","summary":{"takeaway":"Abbasloo and Cheng argue that knowing when to hold an answer and when to yield is its own skill, and that a model can train it by arguing against itself.","found":"The authors say that standard tests ask a model one question alone, while real use often involves people contesting answers. They name the skill of holding a correct position or yielding to a better one \"dialectical competence.\" They report that it depends on first-answer accuracy but does not follow from it. A model can get better at answering without getting better at this, and training for accuracy alone can work against it. They also say it cannot be gained at inference time by self-reflection or by sampling more. Their method, Self-Disagreement, has the model state a position, write the strongest case against it, and rule on whether that case stands. It is then trained on the ruling with reinforcement learning from verifiable rewards. The abstract says the skill is weak in the base model, arrives with the reward, does not cost first-answer accuracy, and holds against challengers the model never trained against, including stronger ones. The abstract gives no numbers, model names or benchmarks.","matters":"Models that drop correct answers when someone pushes back with bad reasoning are hard to rely on. This paper treats that as a separate weakness from accuracy and proposes a way to train against it. The training needs no second model, no reward model, no preference data and no human annotation, so it could be cheap to apply. The abstract does not say how large the gains are, so the strength of the claim is unclear from it alone.","race":"If the result holds, builders could train models to stand firm on correct answers using only checkable rewards and no extra models or human labels. That would lower the cost of a robustness skill, but the abstract gives no figures, so the size of the effect is unknown.","safety":"The work bears on reliability under pressure. A model that yields to wrong counter-arguments can be talked out of correct answers, and one that yields to better ones stays open to correction. The abstract does not test misuse or adversarial settings beyond challengers.","level":"intermediate","tags":["alignment","reinforcement-learning","language-models","evals","training","reasoning"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Soheil Abbasloo, Peng Cheng

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 18, 2026 · Number 4Optimistic Verifiers Get Hacked: Reward Hacking Under an Exact Process OracleWhen the checker used to reward an SI model is too generous, the model learns to game it. The authors show this clearly because they can measure the true reward exactly.Reinforcement learningAlignmentEvals](https://thesimind.si/papers/2611.21538/) · [Nov 9, 2026 · Number 6Alignment by Confident Debate: Preference Learning from Multi-Agent DebateAbCD turns debates among several SI models into training preferences without gold answer labels, and the authors report better accuracy than ensembling, debate alone, or SFT on the same traces.AlignmentReasoningMulti-agent systems](https://thesimind.si/papers/2611.09568/) · [Nov 16, 2026 · Number 6Alignment with Reflect-Reasoning: Learning How To Follow PrinciplesReflect-Reasoning trains models to reason about whether they are following a given set of principles, and the authors report it beats system prompting across five models.AlignmentReasoningLanguage models](https://thesimind.si/papers/2611.17398/)

[Number 11](https://thesimind.si/papers/2611.17461/) · [All 12 for Nov 16, 2026](https://thesimind.si/papers/day/2026-11-16/)
