Recognition–Refusal Misalignment in LLMs: A Routing-Failure Account of Structurally Unanswerable Questions

The abstract of arXiv:2608.29109 reports that instruction-tuned models from 1.7B to 70B parameters represent structural impossibility in a linear hidden-state direction that separates answerable from structurally impossible math and code prompts. That recognition direction is nearly orthogonal to the canonical safety-refusal direction, and an in-domain behavior-defined invalidity-aware direction is only partially aligned with it. Generation-time steering along the recognition direction changes invalidity-aware behavior bidirectionally and dose-responsively on structural math and code cells, while random directions do not; base/instruct comparisons place the low-cosine geometry at the pretraining endpoint. The paper concludes that answering structurally unanswerable questions is better explained as a routing failure than an encoding failure.

Hugging Face Daily Papers — Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions

September 10, 20261 source40% confidence

66Strategic

Strategic score

Evidence strength45

Technical novelty82

Commercial potential55

New: the abstract reports that across instruction-tuned models from 1.7B to 70B parameters, a single linear direction in the hidden state separates answerable from structurally impossible math and code prompts. The same direction is nearly orthogonal to the canonical safety-refusal direction that mediates trained harmful-content refusal. The paper frames the confident-on-impossible failure as a routing failure rather than an encoding failure: the model carries a usable "no admissible answer" signal, but the safety-refusal pathway is not aligned to use it.¹

Context

The constraint that moved is the implicit assumption that safety-refusal alignment is a sufficient gateway for abstention on structurally unanswerable queries. The abstract distinguishes structural invalidity, such as computing cot(-540°) or evaluating (1).startswith("1"), from harmful content and reports that an in-domain behavior-defined invalidity-aware direction is only partially aligned with the recognition direction and near-orthogonal to safety refusal. If that geometry holds, mitigation cannot rely on the existing safety-refusal pathway; it must route an already-present invalidity representation to an abstention decision.¹

Assessment

Evidence-grounded assessment: the causal claim rests on generation-time steering along the recognition direction, which reportedly changes invalidity-aware behavior bidirectionally and dose-responsively on structural math and code cells, while random directions do not. The pretraining-origin claim rests on base/instruct comparisons showing the low-cosine geometry at the pretraining endpoint. These are strong mechanistic assertions, but the abstract supplies no effect sizes, no uncertainty estimates, no prompt sets, no probe methodology, and no list of models beyond the 1.7B–70B range. What would change the conclusion is evidence that linear separation is driven by surface features of the example types rather than by structural impossibility, that steering effects fail to generalize beyond the tested math and code cells, or that the safety-refusal direction can be used for abstention under a different readout.¹

Caveats

Caveats: this is one document, and only the abstract is available in the supplied source. The arXiv record shows no peer-review status, and the visible page contains bibliographic and platform headings rather than code, data, or supplementary materials. The claims about hidden-state geometry, steering causality, and pretraining origin therefore remain abstract-level until the full paper, datasets, and evaluation protocol are inspected.¹

Source text

Hugging Face Daily Papers — Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions · August 29, 2026

Computer Science > Computation and Language

Title:Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions

Abstract:Large language models often answer structurally unanswerable questions, such as computing cot(-540°) or evaluating (1).startswith("1"), instead of abstaining. We ask whether this failure reflects missing recognition or failed routing from recognition to abstention. Across instruction-tuned models from 1.7B to 70B parameters, a single linear direction in the hidden state separates answerable from structurally impossible math and code prompts, showing that models represent impossibility before generation. Yet this recognition direction is nearly orthogonal to the canonical safety-refusal direction that mediates trained harmful-content refusal. An in-domain behavior-defined invalidity-aware direction is closer to recognition, but only partially aligned with it, and remains near-orthogonal to safety refusal. Generation-time steering along the recognition direction changes invalidity-aware behavior bidirectionally and dose-responsively on structural math and code cells, while random directions do not. Base/instruct comparisons further show that the low-cosine geometry is already present at the pretraining endpoint. The confident-on-impossible failure is therefore better explained as a routing failure than as an encoding failure: the model has a usable "no admissible answer" signal, but the safety-refusal pathway is not aligned to use it.

Access Paper:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Unstated assumption: “a single linear direction … separates answerable from structurally impossible math and code prompts” is a linear-separability claim, not proof of an abstract impossibility concept. Separation can arise from surface regularities of the example families, such as the presence of a known impossible pattern, even when the model does not represent structural unanswerability in the intended sense. The abstract does not report probe accuracy, effect size, or controls for lexical and formatting confounds.

Term of art: “canonical safety-refusal direction” refers to a prior construct that mediates trained refusal of harmful content; “nearly orthogonal” means a low absolute cosine between that direction and the recognition direction. Orthogonality is a geometric relation, not evidence of functional independence. Two directions can be near-orthogonal yet both participate in the same downstream decision, so the routing-failure interpretation requires the additional steering and base/instruct evidence.

Claim the evidence does not carry: “the model has a usable ‘no admissible answer’ signal” is an inference from the existence of a separable and steerable direction. Usability requires that the signal be read out by the model's own decision process without external intervention; the abstract shows steering changes behavior, which demonstrates that the direction can be manipulated, not that the model already routes it to abstention. The base/instruct geometry result further shows presence at pretraining, not readiness for use.

The source text above remains the property of its publisher and is reproduced for commentary.

Hugging Face Daily Papers — Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions

August 29, 2026endistinct angle

Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions

↑ Back to top