Lexical Redefinition Reveals Context-Sensitive Instruction Following

Roksana Goworek; Haim Dubossarsky

cs.CLOct 12, 2026

Lexical Redefinition Reveals Context-Sensitive Instruction Following

Authors: Roksana Goworek, Haim Dubossarsky

Organizations: Queen Mary University of London · University of Cambridge

Abstract

How do language models reconcile conflicting contextual evidence within a prompt? We study this question using lexical redefinition, where an explicit instruction specifies how a word should be interpreted and a subsequent sentence provides additional evidence through its usage. Across 870 word pairs, we find that lexical redefinitions are not applied as fixed rules. Instead, the behaviour depends systematically on the relationship between the explicit instruction and the intervening sentence: the contextualised usage can preserve, weaken, or completely reverse the instruction. These results suggest that the interpretation of an explicit instruction remains sensitive to subsequent contextual evidence. Lexical redefinition therefore provides a controlled behavioural paradigm for studying how language models integrate multiple sources of information within a prompt.

Figures & tables

Figure 1: Token probabilities for sentence-completion queries under redefinition prompts and the original prompt without redefinition (" $\varnothing$ "). Sentences and icons are illustrative, results are aggregated over 870 word pairs. Error bars show SD.

Sentence
“The mouse eats crumbs” $\therefore implies$	Compatible uses in its ordinary sense, but the redefinition affects only contexts $\rightarrow$ redefinition unaffected or slightly weakened	Competes uses in its ordinary sense, which was redefined to be denoted by $\rightarrow$ adopted instead
“The letter eats crumbs” $\therefore implies$	Contradicts shows an opposite mapping to the one in the instruction $\rightarrow$ adopted instead	Reinforces shows an example of the redefinition $\rightarrow$ redefinition effect is unchanged

Table 1: Analysis of the four sentence–redefinition combinations. Each cell names the relationship between the sentence and the redefinition instruction, and the observed effect on the completions.

Figure 2: Effect of the intervening sentence on completion probabilities ( $\Delta P$ , Eq. 2 ). Blue/orange bars show $\Delta P$ for mouse / letter . Percentages inside boxes indicate sign agreement with the mean across word pairs (Eq. 3 ).

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Nouns	Verbs	Adjectives
truth	run	hot
chair	know	cold
dog	break	big
flower	give	small
bank	think	good
river	find	bad

Appendix

Table 2: The 30-word stimulus set by part of speech.

Wording	Prompt (X, Y stand for the word pair)
baseline	In the following sentences, the word ‘X’ should be interpreted as if it were the word ‘Y’.
scope_sentence_only	For the next sentence only, ‘X’ should be interpreted as if it were ‘Y’.
scope_global	For the rest of this conversation, ‘X’ should be interpreted as if it were ‘Y’.
denotational	Whenever ‘X’ appears, read it as ‘Y’.
comprehension_only	Treat ‘X’ as having the meaning of ‘Y’ when reading. Do not change how you produce these words; only how you interpret them.

Appendix

Table 3: The baseline redefinition prompt and four alternative formulations.

Figure 3: Qwen2.5-14B-Instruct: raw next-token probabilities. Rows: context sentence (icon) $\times$ query, plus two no-sentence rows; columns: redefinition condition (icon). Bars: P(mouse) $=w_{1}$ , P(letter) $=w_{2}$ ; $\pm 1$ SD over pairs.

		$\varnothing$
context	query
		.10 [1pt] $\pm$ .14	.00 [1pt] $\pm$ .00	.07 [1pt] $\pm$ .20	.01 [1pt] $\pm$ .02	.34 [1pt] $\pm$ .19	-.28 [1pt] $\pm$ .26
		.00 [1pt] $\pm$ .01	.07 [1pt] $\pm$ .10	.05 [1pt] $\pm$ .22	.02 [1pt] $\pm$ .02	.30 [1pt] $\pm$ .23	-.29 [1pt] $\pm$ .25
		.00 [1pt] $\pm$ .00	.10 [1pt] $\pm$ .19	-.36 [1pt] $\pm$ .28	.50 [1pt] $\pm$ .21	.00 [1pt] $\pm$ .01	.16 [1pt] $\pm$ .23
		.04 [1pt] $\pm$ .08	.01 [1pt] $\pm$ .02	-.44 [1pt] $\pm$ .28	.52 [1pt] $\pm$ .27	.00 [1pt] $\pm$ .02	.15 [1pt] $\pm$ .28

Appendix

Table 4: No-sentence-normalised sentence effect

\Delta P

(

P_{\text{with sentence}}-P_{\text{no sentence}}

) for Qwen2.5-14B-Instruct under the scope_global wording (8-word variant subset). Rows: context sentence

\times

query word; columns: redefinition direction

\times

queried word. Each cell: mean

\Delta P

over the 56 pairs;

\pm

line is the SD over pairs.

		$\varnothing$
context	query
		.10 [1pt] $\pm$ .14	.00 [1pt] $\pm$ .00	-.15 [1pt] $\pm$ .16	.02 [1pt] $\pm$ .03	.20 [1pt] $\pm$ .15	-.21 [1pt] $\pm$ .21
		.00 [1pt] $\pm$ .01	.07 [1pt] $\pm$ .10	-.08 [1pt] $\pm$ .18	.02 [1pt] $\pm$ .03	.11 [1pt] $\pm$ .15	-.19 [1pt] $\pm$ .22
		.00 [1pt] $\pm$ .00	.10 [1pt] $\pm$ .19	-.30 [1pt] $\pm$ .23	.29 [1pt] $\pm$ .18	.00 [1pt] $\pm$ .01	-.10 [1pt] $\pm$ .22
		.04 [1pt] $\pm$ .08	.01 [1pt] $\pm$ .02	-.35 [1pt] $\pm$ .21	.21 [1pt] $\pm$ .17	.02 [1pt] $\pm$ .02	-.14 [1pt] $\pm$ .20

Appendix

Table 5: Prompt-variation results for Qwen2.5-14B-Instruct ( scope_sentence_only wording).

		$\varnothing$
context	query
		.10 [1pt] $\pm$ .14	.00 [1pt] $\pm$ .00	-.02 [1pt] $\pm$ .24	-.09 [1pt] $\pm$ .18	.22 [1pt] $\pm$ .23	-.13 [1pt] $\pm$ .18
		.00 [1pt] $\pm$ .01	.07 [1pt] $\pm$ .10	-.04 [1pt] $\pm$ .13	-.01 [1pt] $\pm$ .21	.25 [1pt] $\pm$ .16	-.20 [1pt] $\pm$ .23
		.00 [1pt] $\pm$ .00	.10 [1pt] $\pm$ .19	-.15 [1pt] $\pm$ .19	.24 [1pt] $\pm$ .26	-.13 [1pt] $\pm$ .17	.11 [1pt] $\pm$ .27
		.04 [1pt] $\pm$ .08	.01 [1pt] $\pm$ .02	-.24 [1pt] $\pm$ .24	.26 [1pt] $\pm$ .17	-.07 [1pt] $\pm$ .21	.05 [1pt] $\pm$ .16

Appendix

Table 6: Prompt-variation results for Qwen2.5-14B-Instruct ( denotational wording).

		$\varnothing$
context	query
		.10 [1pt] $\pm$ .14	.00 [1pt] $\pm$ .00	.01 [1pt] $\pm$ .24	.00 [1pt] $\pm$ .01	.12 [1pt] $\pm$ .10	-.17 [1pt] $\pm$ .16
		.00 [1pt] $\pm$ .01	.07 [1pt] $\pm$ .10	.01 [1pt] $\pm$ .23	.00 [1pt] $\pm$ .01	.13 [1pt] $\pm$ .12	-.12 [1pt] $\pm$ .22
		.00 [1pt] $\pm$ .00	.10 [1pt] $\pm$ .19	-.33 [1pt] $\pm$ .24	.36 [1pt] $\pm$ .16	.00 [1pt] $\pm$ .01	.08 [1pt] $\pm$ .29
		.04 [1pt] $\pm$ .08	.01 [1pt] $\pm$ .02	-.38 [1pt] $\pm$ .19	.35 [1pt] $\pm$ .18	.00 [1pt] $\pm$ .01	.00 [1pt] $\pm$ .26

Appendix

Table 7: Prompt-variation results for Qwen2.5-14B-Instruct ( comprehension_only wording).

LLM Prompting

Large language model prompting studies how instructions, examples, context, and formats steer model behavior at inference time.

See all 419 papers

May 19, 2026

Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs

Language models are trained to follow instructions, but they are also powerful pattern completers. What happens when these two objectives conflict? We construct conversations in which a user instruction to behave in a target way T (e.g., always output a specific token, answer in a particular language, or adopt a persona) is opposed by N hardcoded assistant turns demonstrating a competing pattern P. We then measure instruction-following (IF) rates in this setting, across 13 models and 16 different instructions, for up to 50 turns. Average instruction-following rates range from 1% to 99% across models, largely uncorrelated with standard capability benchmarks. The transition from instruction-following to pattern-following is universal but highly model-dependent. Robustness is modulated both by instruction content, with models resisting induction longer when instructions align with their trained value priors, and by output format, with diverse multi-token responses proving substantially more resistant than single-token outputs. Chain-of-thought reasoning improves robustness but does not eliminate susceptibility, and can produce dissociation between correct deliberation and incorrect output. When asked to predict their behavior in this setting, models achieve 83.5% accuracy on average but systematically underestimate their own resistance to induction pressure. These results suggest that instruction-following remains brittle under induction pressure even for otherwise capable models, and that output diversity, rather than semantic engagement with the input, is the primary factor predicting robustness.

Carolina Camassa, Derek Shiller

Future Impact Group · Rethink Priorities

LLM Evaluation Language Model Self-Assessment Instruction Following In-Context Learning Language Model Robustness

Ask PDF

May 25, 2026

Priors Persist Through Suppression: A Stroop Paradigm for Lexical Override

Glossaries, technical specifications, and system prompts routinely ask language models to use familiar words in unfamiliar ways. When this works, the local rule does not overwrite the old meaning; the pretrained prior keeps operating underneath, and its strength still shows through. We test this with a Stroop-style paradigm: a remapping rule (doctor means forest) pitted against the query word's lexical-prior distractor (hospital), with matched neutral controls. Across 11 open-weight models spanning four families and 1B-9B parameters, lexical-prior strength predicts interference even after item-level controls for answer prior, frequency, tokenization, and prompt wording. Activation patching on five models then locates where the override is repaired internally. Restoring three source positions that carry the redefinition (the definition subject, its new target, and the query word) almost fully recovers the effect (aggregate

R \in [0.92, 1.06]

). The repair works by protecting the contextual target rather than by silencing the prior; the distractor's probability falls whenever these positions are perturbed, but the target survives only when the redefinition is restored intact. Behavior and mechanism converge on the same channel: the prior's strength both predicts which overrides fail and marks where the causal repair lands.

Han-yu Wang

The University of Hong Kong

Knowledge Conflicts in Language Models LLM Interpretability Causal Interventions in Language Models

Ask PDF

Jul 21, 2026

Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models

Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade. We report two controlled experiments crossing all three factors on one held, contamination-free synthetic corpus (the "Book of Veyra," 8,780 uniquely-named entities, deterministically regenerable from a fixed seed), evaluated across five models. Experiment 1 (960 calls/model) measures instruction-following decay as rule count N grows from 10 to 160, crossed with four formats and system-prompt vs. user-turn placement. Perfect-response rate collapses to zero by N=80 for every model, format, and placement. Placement produces effects at least as large as format at N=160 in most models, but the direction is model-specific. No model shows a reliable markdown advantage; one 35B model favors plain text instead. Experiment 2 (5,520 calls/model) measures recall accuracy, false-premise sycophancy, and absent-fact fabrication across a 2k-to-512k-token context ladder in the same four formats. Recall stays near ceiling through 64-128k tokens, then degrades sharply and format-dependently: one model's accuracy spread reaches 48 points at 128k tokens. Fabrication never occurs (0/5,760 probes), and sycophancy stays negligible (<=8.3%). What rises sharply near each model's context ceiling is outright refusal to answer (0% to 79-90%), distinct from sycophancy or fabrication. Neither pre-registered format ordering holds, and token overhead (+22% to +37% over plain text) further changes which format is preferable where accuracy spread is genuine. We release the full harness, corpus generator, and raw results (VeyraBench): https://github.com/iNetanel/veyrabench

Netanel Eliav

Machine Human Intelligence Lab (MHIL)

LLM Evaluation Language Model Generation Evaluation Long-Context Language Model Evaluation LLM Prompting Hallucination in Language Models Instruction Following+1

Ask PDF

Abstract

LLM Prompting

Explore similar work

Do as I Say, Not as I Do: Instruction-Induction Conflict in LLMs

Priors Persist Through Suppression: A Stroop Paradigm for Lexical Override

Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models