Lexicality Heuristic in Language Models

• Lexicality heuristic is the tendency of models to use surface-level word overlaps instead of integrating deeper context for inference.
• Experimental methodologies quantify this heuristic using step-wise bias ratios and conditional probabilities, emphasizing its decline as reasoning progresses.
• Mitigation strategies such as dataset filtering and adversarial approaches are proposed to reduce spurious lexical biases and improve model robustness.

The lexicality heuristic refers to the systematic tendency of machine learning models, especially LLMs and natural language inference (NLI) systems, to rely on superficial lexical cues—such as overlapping words or specific tokens—rather than meaningfully integrating contextual or logical relationships. This heuristic can drive both model predictions and multi-step inference trajectories, often leading to correct answers for the wrong reasons or errors in settings designed to require deeper reasoning. Rigorous experimental work has quantitatively isolated, measured, and analyzed the lexicality heuristic in multiple domains, notably multi-step reasoning with LLMs and hypothesis-only approaches in NLI.

1. Formal Definition and Taxonomy

The lexicality heuristic encompasses any strategy in which a model exploits surface-level lexical overlap or word-label co-occurrences, rather than engaging in goal-directed or context-integrative inference. A key instantiation is preference for premises that share surface strings, most notably proper names, with the target question, even when such overlap is logically irrelevant. In the context of NLI, the related notion of lexical bias is operationalized as a deviation from a uniform label distribution over hypotheses conditioned on the occurrence of a given word or property—i.e., $P(L|w)$ being strongly skewed for particular $w$.

In multi-step reasoning tasks, the overlap (or lexicality) heuristic is often instantiated as follows:

• For a given question $q$ and a set of premises $P = \{p_1, \ldots, p_K\}$, a distractor premise $\tilde p$ is designated as overlapping if it contains exactly the same person name (PN) as the question, irrespective of logical relevance (Aoki et al., 2024).
• In NLI, any word $w$ or higher-level property $pr$ creates lexical bias if the conditional label distribution $P(L|w)$ or $P(L|pr)$ is substantially non-uniform—enabling label prediction prior to any reference to the premise (Hu et al., 2021).

2. Experimental Methodologies

Multi-Step Reasoning Framework

The operationalization in (Aoki et al., 2024) involves constructing controlled reasoning tasks. For example, in arithmetic reasoning, an irrelevant premise (the “Base” distractor) is surgically rewritten so that its subject matches the question’s person name, thereby ensuring lexical overlap. This construction enables direct measurement of the model’s inclination to select $\tilde p$ at different reasoning steps.

• Step-wise bias ratio: For a minimal solution $\bm{h}^*$ with $t$ steps, at each stage $d$ (distance to goal), define

$$r(d) = \frac{\#\{\text{model picks } \tilde p_t \text{ when } g(\bm{h}^*_{\leq t})=d\}} {\#\{\text{model picks correct } h^*_t \text{ when } g(\bm{h}^*_{\leq t})=d\}}$$

where $r = 0.5$ reflects random choice, and $r > 0.5$ quantifies bias toward lexical distractors.

NLI Lexical Bias Detection

In NLI, (Hu et al., 2021) quantifies lexical bias using word-level and proto-role co-occurrence statistics:

• For each word $w$ in hypotheses $H$, $P(L|w) = \text{count}(w, L) / \text{count}(w)$ is computed.
• Proto-role majority bias for property $pr$:

$$\mathrm{maj}_{pr} = \frac{\text{count}(l_{pr}, pr)}{\text{count}(pr)}$$

where $l_{pr}$ is the majority label for $pr$.

3. Quantitative Findings

Experimental evidence highlights several consistent patterns:

Model or Setting	Overlap/Heuristic Bias	Key Numerical Results
Multi-Step Arithmetic (Aoki et al., 2024)	r(d) declines with step index	PaLM2: r(4) ≈ 0.9 → r(1) ≈ 0.3
	Single-step Overlap bias	PaLM2: Base 10.3%, Overlap 42.3%
		Llama2: Base 32.6%, Overlap 67.7%
NLI (SPR dataset) (Hu et al., 2021)	Proto-role bias	“Stationary”: maj=0.96+
	Word-level bias	“that”: P(not-entailed

• In multi-step tasks, models are substantially more likely to select a distractor premise sharing the question PN at early reasoning steps ($d$ large). This bias decreases as the model approaches the final answer, implying that the lexicality heuristic is more influential when the answer is distant and is “abandoned” in favor of goal-directed steps as the solution progresses.
• In NLI, high-frequency proto-roles or words such as “stationary,” “that,” and “market” display extreme label skews. For instance, “stationary” nearly always predicts the same class, conferring $>96\%$ accuracy if used as a classifier; likewise, the function word “that” strongly signals “not-entailed.”

4. Theoretical Interpretation and Model Behavior

Findings from (Aoki et al., 2024) support the hypothesis that LLMs employ a bounded-horizon search. When the solution is remote, models exploit shallow lexical heuristics, such as overlap, to prune the search space or expedite reasoning. As the reasoning chain approaches the answer (i.e., as $d \rightarrow 1$), models transition to more rational, logically-driven computation.

This two-stage behavior echoes classic theories of human problem-solving, where initial steps are guided by heuristics and later refined by analytic reasoning. The “first-macro, finishing-rational” blend helps explain the performance improvements from chain-of-thought prompting (which extends rational lookahead) and also why LMs often fail on complex, long-horizon problems—early heuristic choices propagate errors (Aoki et al., 2024).

In NLI, the lexicality heuristic undermines the intended semantic comparison between premise and hypothesis. Models can achieve high accuracy on datasets like SPR by guessing labels from surface statistics in the hypothesis alone, revealing a persistent confound in many widely used benchmarks (Hu et al., 2021).

5. Statistical Measures for Detection

The detection and quantification of lexicality reliance leverage:

• Conditional probabilities: $P(L|w)$ and $P(L|pr)$ computed from gold-labeled corpora.
• Chi-square independence tests: Assessment of whether role properties and labels are statistically independent yields extremely significant results (e.g., $\chi^2=30\,632$, $p<2.2 \times 10^{-16}$) (Hu et al., 2021).
• Step-wise bias ratios: Measurement of how $r(d)$ decays as $d$ decreases (number of reasoning steps remaining) gives a precise profile of heuristic reliance in LLMs (Aoki et al., 2024).

6. Mitigation Strategies

Several approaches are proposed to counteract the lexicality heuristic:

• Dataset filtering: Remove “give-away” tokens or roles with extreme label bias (e.g., high $\mathrm{maj}_{pr}$ or $P(L|w)$).
• Adversarial filtering: Dynamically mine or generate instances that break spurious lexical correlations.
• Stricter benchmarks: Require that premise-aware models significantly outperform hypothesis-only or heuristic-based baselines.
• Counter-balanced splits: Ensure each lexical unit (word or property) appears equally often with both labels.
• Regularization/debiasing: Objective functions or data transformations that penalize model reliance on single tokens.

These techniques aim to ensure that models must engage in genuine inference rather than achieving high accuracy via distributional artifacts (Hu et al., 2021).

7. Implications for Model Evaluation and Development

Rigorous isolation and measurement of the lexicality heuristic provide a blueprint for understanding and mitigating shortcut reasoning in current models. In multi-step reasoning, quantitative tracking of heuristic usage (e.g., through $r(d)$) illuminates both LM strengths (flexible, dynamic strategies) and their fundamental limitations (shallow horizon, susceptibility to superficial cues). In NLI, documenting lexical biases is critical for trustworthy evaluation of “understanding”-oriented systems and for the future design of robust, artifact-resistant data and architectures (Hu et al., 2021, Aoki et al., 2024).

A plausible implication is that careful engineering of benchmarks and training procedures will be required to eliminate right-for-the-wrong-reasons solutions and foster genuine semantic or logical competence in LLMs.