QK-score attention alignment as a lightweight alternative to ablation for testing logical consistency

From ablation to attention alignment: what earlier evaluation established

Before QK-score, the dominant way to test whether a model actually used a piece of reasoning or evidence was to remove it and observe the effect. In fact-checking, multi-round evidence ablation showed that many LLM verifiers keep their predictions even after supporting evidence is removed, revealing reliance on parametric knowledge rather than retrieved evidence [4]. That work also made the cost explicit: iterative evidence ablation and repeated verifier inference make evaluation computationally heavier than standard single-pass fact-checking, so it is better suited to diagnosis than large-scale lightweight evaluation [4]. Separately, statistical analysis of chain-of-thought prompting modeled the estimator induced by CoT as approximating a Bayesian estimator that solves the reasoning task by inferring latent steps, and proved that longer reasoning chains reduce prediction error in expectation while raising pretraining costs [3]. Together these established both the value and the expense of probing reasoning through behavior or perturbation.

QK-score enters this lineage by avoiding perturbation altogether. Instead of ablating components or evidence, it computes a single forward pass and extracts a score from the natural alignment between a specific query and the key vectors of the last statement token inside chosen attention heads [1]. The paper positions this as a complement to existing techniques and an efficient alternative to ablation-based analysis [1]. The interpretation is that logical consistency may be readable from internal alignment patterns rather than only from output changes after intervention, but that reading depends on selecting the right heads.

What QK-score measures and how the heads are chosen

The method relies on multiplying query and key vectors and comparing elements of the resulting matrix corresponding to specific tokens, which the authors note aligns directly with the core operations of transformer inference [1]. Heads are selected using a calibration set: the paper states the method requires a sufficiently large calibration dataset of at least roughly 400 reasoning questions, with balanced coverage of logical rules and careful debiasing to avoid selecting heads that exploit question phrasing rather than logic [1]. In the reported experiments, strongly pronounced logic heads typically emerge after the first third of layers, with Qwen models concentrating them in the final half or final third and LLaMA models populating the latter two-thirds; these heads alternate with inverse-logic heads that consistently predict the wrong answer [1].

The evaluation spans logical reasoning benchmarks including ProntoQA-OOD, PARARULE Plus, and an extended Multi-LogiEval, across models from 1.5B to 70B parameters [1]. On ProntoQA-OOD, QK-score often exceeds the model's own final-layer baseline, and the gap widens with reasoning depth and distractors. For example, LLaMA3.1 8B Instruct reaches QK-score accuracy of 0.9881 at depth 1 and 0.8676 at depth 5 on Modus Ponens, versus baseline 0.9843 and 0.8775, while on composition of rules the same model scores 0.8084 at depth 1 and 0.7052 at depth 5 against baseline 0.6848 and 0.5731 [1]. The pattern is not uniform: LLaMA3.1 8B Base shows QK-score 0.8909 at depth 1 falling to 0.6851 at depth 5, below its baseline at deeper hops, which the authors do not present as a universal win [1].

Distractors, reasoning depth, and cross-domain behavior

The paper's robustness claim rests on two stress tests. With distractors added to Modus Ponens prompts, QK-score generally degrades more slowly than baseline. LLaMA3 8B Instruct drops from 0.9611 with one distractor to 0.9220 with five, while its baseline falls from 0.9474 to 0.8361; Qwen-2.5 32B Instruct goes from 0.9997 to 0.9429 against baseline 0.9990 to 0.7522 [1]. On reasoning depth, the same pattern holds for several models: Qwen-2.5 14B Instruct scores 0.9988 at depth 1 and 0.9683 at depth 5 on Modus Ponens versus baseline 0.9660 and 0.9060, and 0.9724 to 0.8397 on composition versus baseline 0.9050 to 0.7451 [1]. These numbers indicate that the selected heads carry a signal that survives surface changes in the prompt, at least within ProntoQA-OOD.

Cross-domain transfer is more mixed. On PARARULE Plus, three of five heads selected on ProntoQA-OOD exceed the baseline, in most cases by more than 10%, but head (22,16) from DeepSeek-R1 consistently yields accuracy below 0.35, suggesting its QK-score distinguishes correct from incorrect implications in a reversed manner; a similar reversal occurs in some setups on other heads [1]. The authors also report that mixing ProntoQA-OOD and FOLIO for head selection yields better average accuracy on the target dataset than either alone, while heads (22,26) and (18,1) remain in the top five across cases, which they read as support for the claim that the method selects heads responsible for logic processing [1]. The reversal is a genuine boundary: a high QK-score is not universally a validity signal without checking head orientation.

Ablation comparison and what the score does not capture

The paper directly compares QK-score with ablation by pruning the K best attention heads and comparing against pruning K random heads on ProntoQA-OOD for LLaMA-3.1-8B with K = 10 and 20, averaged over seven restarts [1]. This is a narrower ablation test than the multi-round evidence ablation used in fact-checking, which repeatedly removes a model's own predicted evidence to characterize causal dependency between selected evidence and verdict [4]. The two designs answer different questions: QK-score asks whether internal alignment predicts logical validity, while evidence ablation asks whether a model's stated evidence causally supports its verdict. The fact-checking work also found that strong initial accuracy alone may not reflect genuine evidence dependency, since several baselines maintained high verification performance after substantial evidence ablation [4].

The paper is explicit about what QK-score does not incorporate. It does not directly use value vectors or output aggregation matrices, which are also important components of attention, and the authors state that studying those components requires different approaches outside the paper's scope [1]. It also locates individual heads rather than head combinations, and the authors caution that no single head will outperform all others on every logical task, that dataset-specific biases affect the best-performing heads, and that the identified heads are not claimed to be the only ones responsible for logical inference [1]. Head selection was conducted on only two sets of deduction rules, Modus Ponens alone and a broader set including conjunction/disjunction introduction/elimination and proof by contradiction, so heads specialized for other rules may have been missed [1].

Boundaries of the claim and open questions

The evidence boundary is set by the tested models and tasks: conclusions are limited to the 1.5B to 70B models and logical reasoning benchmarks used, and generalization to other tasks or architectures is not demonstrated [1]. The calibration requirement of at least about 400 balanced, debiased reasoning questions is a practical constraint that may be difficult to satisfy for specialized domains [1]. The paper also does not establish that QK-score captures faithfulness in the sense used by newer chain-of-thought benchmarks, which distinguish post-hoc rationalization from spurious reasoning chains and find only weak association between correctness and faithfulness, with a phi correlation of 0.286 and mutual information of 0.057 in one annotated dataset [6]. A high QK-score could in principle reflect a coherent-looking but unfaithful trajectory, and the supplied evidence does not test that.

Broader interpretability work offers context rather than validation. A systematic review of attention interpretation describes probing methods as typically combining a probing task with a probing classifier, which is a different methodological family from directly reading query-key alignment [5]. Foundational work on transformer modularity and on attention in medical imaging shows that attention heads can be regularized with anatomical and spatial priors and that attention modules can be decomposed and reused, but these do not test logical consistency in language models [2][7]. The priority-map analogy from visual neuroscience proposes that both brains and language models dynamically prioritize information through bottom-up and top-down signals, which is a conceptual parallel rather than evidence about QK-score [8]. Whether QK-score heads correspond to any stable reasoning circuit, and whether they can be combined with chain-of-thought prompting as the authors suggest for future work, remains untested [1].

About These Sources

This research page is built on 8 peer-reviewed studies — published from 2024 to 2026, 8 from 2024 or later, collectively cited 71 times — selected as the most relevant from 13 studies that passed quality screening, drawn from 96 papers retrieved from a database of over 500 million.

Sources used in this answer