Source-linked AI summary
The Reliance Hypothesis: Chain-of-Thought Faithfulness Rises as Language Resource Level Falls
Lucas Resck, Jingyi Sun, Qianli Wang, Isabelle Augenstein, Anna Korhonen
TL;DR
CoT faithfulness matters for monitoring and safety, but research has been conducted almost entirely in English, leaving its relationship to language resource level and task type underexplored. This paper relates faithfulness to resource level across task types and finds it rises as resource level falls on knowledge-intensive tasks, but not reasoning-intensive ones.
Problem
CoT faithfulness matters for monitoring and safety, but research has been conducted almost entirely in English, leaving its relationship to language resource level and task type underexplored.
Source: §1 · pp. 1–2Method
The study relates resource level to faithfulness across task types, estimating separate slopes and testing them with Fisher randomisation.
Source: §3.2 · p. 3Results
Faithfulness rises as resource level falls on knowledge-intensive tasks, but not reasoning-intensive tasks; this gap is broadly consistent across models, metrics and robustness checks.
Source: §5 · p. 8Takeaways & Limitations
CoT monitoring may be most informative where models know least and CoT faithfulness is least often validated.
Source: §5 · p. 8Only eight languages are studied, with Swahili and Yoruba carrying most resource-level spread, so estimates may not generalise to lower-resource languages broadly.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Chain-of-Thought (CoT) faithfulness matters for monitoring, explaining and safely deploying large language models, yet it has been studied mostly in English. Languages differ in level, the amount of text available to learn from, so models know less in lower-resource languages, which may change how much they rely on their reasoning. To our knowledge, we are the first to relate CoT faithfulness to both resource level and task type (knowledge- or reasoning-intensive). On parallel multiple-choice questions from Multilingual Massive Multitask Language Understanding (MMMLU) in eight languages, faithfulness rises as resource level falls on knowledge-intensive tasks but not on reasoning-intensive ones, a gap that points the same way across four models and several faithfulness measures. We propose, and find support for, the Reliance Hypothesis: models rely more on their reasoning when they cannot recall the answer, as in lower-resource languages on knowledge-intensive tasks. Model confidence partly explains the effect of resource level, but not the knowledge-reasoning gap. Languages in which models are less confident rely more on reasoning, yet less confident answers within a language do not, a Simpson's paradox. In lower-resource languages, models rely on their reasoning for more answers, not more strongly for each. CoT monitoring may therefore be most informative in lower-resource languages, where it is least often validated.
1 Introduction
The introduction frames multilingual CoT faithfulness as relevant to monitoring, explanation, and safety, while noting that its relationship to language resource level and task type remains open. It tests that relationship and proposes the Reliance Hypothesis: models lean more on reasoning when they cannot recall an answer.
1 Introduction
- The Reliance Hypothesis proposes that models rely more on reasoning when they cannot simply recall an answer, as in lower-resource languages on knowledge-intensive tasks.The introduction illustrates this with a corrupted chain: a model with thinner knowledge follows the mistake rather than recalling the answer.Source: Figure 1 · p. 1, §1 · p. 2
- The study tests this account on parallel MMMLU questions across eight languages, four open-weight models, and four reasoning-chain interventions, comparing knowledge- and reasoning-intensive subtasks.Faithfulness is measured by how much each intervention shifts confidence in the original answer; the same languages and models serve as controls across task types.Source: §1 · p. 2
- The authors identify relating CoT faithfulness to both resource level and task type as a contribution, addressing a relationship left open by prior multilingual studies.Prior work had reported language differences in faithfulness but had not related them to resource level or contrasted knowledge- and reasoning-intensive tasks.Source: §1 · pp. 1–2
- Faithfulness rises as resource level falls on knowledge-intensive tasks, but not on reasoning-intensive tasks, a pattern consistent across models, metrics, and robustness checks.The introduction reports that the gap also largely survives robustness checks.Source: §1 · p. 2
- Model confidence partly explains the resource-level effect, but does not explain why that effect is stronger on knowledge-intensive tasks.Less confident languages show more reliance on reasoning, yet less confident answers within a language do not; lower-resource languages rely on reasoning more often, not more heavily per answer.Source: §1 · p. 2
- CoT monitoring may be most informative in lower-resource languages, where it is least often validated.The authors also caution that faithfulness is highest where models know least.Source: §1 · p. 2
2 Related Work
Prior work has developed several ways to assess CoT faithfulness, but the measure remains unsettled. Multilingual studies have found language-related patterns without examining their relationship to resource level or contrasting knowledge- and reasoning-intensive tasks.
2 Related Work
- Faithfulness measures remain unsettled without ground truth; approaches intervene on inputs or chains, unlearn stated knowledge, compare attributions, or inspect internal computations.The authors adopt four chain-removal or corruption interventions and three metrics (Lanham et al., 2023).Source: §2 · p. 2
- Multilingual studies report differing patterns, including stronger trace reliance outside English on multilingual maths, while adversarial hints often go unmentioned (Onyame et al., 2026).The findings need not conflict because a chain can drive an answer without mentioning a hint.Source: §2 · p. 2
- Prior multilingual studies had not related faithfulness to resource level or compared knowledge- and reasoning-intensive tasks; this paper derives the Reliance Hypothesis from its observed pattern.The hypothesis parallels evidence that external context helps most on facts models rarely memorise.Source: §2 · p. 2
3 Cross-lingual Faithfulness Variation
Across languages, CoT faithfulness rises as resource level falls on knowledge-intensive tasks but not on reasoning-intensive tasks. The section tests this pattern across metrics, models, and subtasks.
3.1 Experimental Setup
- The setup pairs shared questions across eight languages and eight subtasks with four open-weight models, using four chain interventions.The subtasks are split evenly between knowledge-intensive and reasoning-intensive subjects.Source: §3.1 · p. 3
3.2 Faithfulness Varies Across Languages
- On knowledge tasks, faithfulness rises as resources fall under all but Early Answering; reasoning slopes are never significantly negative, and the knowledge–reasoning gap is significant for three interventions.Under Adding Mistakes, βK = −0.16 predicts a log-probability drop 1.3 larger in Yoruba than English, close to the observed gap of 1.2.Source: Figure 2 · p. 4, §3.2 · p. 4
3.3 Different Metrics, Same Story
- 0.67 is the minimum pairwise Pearson correlation between faithfulness metrics for every intervention; ∆p and ∆log p have Spearman correlations of 0.92–1.00.The two continuous metrics rank generations almost identically.Source: §3.3 · p. 4
3.4 Models Agree on Faithfulness Variation
- Model pairs agree positively (0.10–0.52), and all four models show a lower resource-faithfulness correlation on knowledge than reasoning subtasks.The knowledge–reasoning gap ranges from −0.56 for Aya to −0.09 for Qwen.Source: §3.4 · pp. 4–5, Figure 3 · p. 5
3.5 A Knowledge–Reasoning Gap
- All four knowledge subtasks rank first by resource-faithfulness correlation; their trends are mostly negative, unlike the weaker, less consistent reasoning trends.For valid answers in both English and Yoruba, Adding Mistakes changed answers in Yoruba alone versus English alone 80 vs. 9 times on knowledge tasks, compared with 73 vs. 40 on reasoning tasks.Source: Figure 3 · p. 5, §3.5 · p. 5
4 The Reliance Hypothesis
The Reliance Hypothesis predicts that models lean more on reasoning when they cannot recall an answer. The reported analyses find a larger resource-level faithfulness gradient on knowledge-intensive than reasoning-intensive tasks, while confidence and reliance patterns clarify—but do not fully explain—the difference.
4 The Reliance Hypothesis
- The Reliance Hypothesis predicts greater reasoning reliance when relevant parametric knowledge is unavailable for recall, while reasoning is needed across languages for computed answers.The paper tests whether the knowledge–reasoning difference is stronger on knowledge-intensive tasks and whether confidence mediates it.Source: §4 · p. 5
4.1 Does the Hypothesis Hold?
- The knowledge-minus-reasoning resource slope is negative in 11 of 12 combinations and significant in seven; under ∆log p, it ranges from −0.17 to −0.31 per token-count order.The knowledge slope is significantly negative under every metric, while the reasoning slope stays near zero, most clearly under Adding Mistakes.Source: §4.1 · p. 6
4.2 Does Confidence Explain the Variation?
- Confidence partly accounts for resource-level slopes, but similar confidence gradients across task families cannot explain their faithfulness gap.On knowledge subtasks, final-answer probability and log-probability absorb 87% and 70%; confidence-based attenuations are imprecise because they rely on only eight languages.Source: §4.2 · p. 6
4.3 Is It a Simpson’s Paradox?
- Across languages, mean |r| is 0.47 versus 0.06 within languages; within-language links mostly vanish, making the confidence–faithfulness association a language-level, not item-level, pattern.In the Figure 4 example, the between-language association is β = −10.0, r = −0.97; within-language it is mean β ≈ 0, r = −0.01.Source: §4.3 · p. 7, Figure 4 · p. 7
4.4 Reliance is Rate, not Magnitude
- On knowledge tasks, reliance rises from 4% of items in English to 35% in Yoruba because interventions move answers more often, not more strongly.Across interventions, the knowledge–reasoning difference is significant in the rate but not magnitude for all but Early Answering.Source: §4.4 · p. 7, Figure 5 · p. 8
4.5 Robustness Tests
- The gap stays negative and significant under every added control; across subsets it keeps its sign but shrinks to a quarter to half its primary size, losing significance in the last three.Across interventions and metrics, Filler Tokens is most robust, significant in 19 of 21 specifications.Source: §4.5 · p. 8
5 Discussion and Conclusion
The findings are consistent with the Reliance Hypothesis, but do not establish its mechanism. They suggest that faithfulness varies with language resource level on knowledge-intensive tasks, with implications for multilingual monitoring and evaluation.
Findings and Mechanism
- The results fit the Reliance Hypothesis: lower-resource languages show higher faithfulness on knowledge tasks, but the evidence does not establish the mechanism.Confidence partly explains variation, yet similar confidence gradients across task families do not explain their gap.Source: §5 · p. 8
Implications
- CoT monitoring may be most informative in lower-resource languages, where it is least often validated; English-only evaluations may understate reasoning reliance on knowledge tasks.The paper frames equally reliable monitoring and explanation across languages as a goal.Source: §5 · p. 8
- Reliance is not verbalization: a chain can drive an answer while omitting what influenced it, so both properties need testing across languages.Source: §5 · p. 8
Limitations
The findings are bounded by the language sample and resource-level proxy, translated and culturally skewed data, generation validity, measurement choices, and model size.
Languages
- Only two languages, Swahili and Yoruba, account for most resource-level variation, limiting generalisation and leaving language-family, script, and tokenisation effects entangled.The authors note that more languages, especially between Bengali and Yoruba, would sharpen estimates.Source: Limitations · p. 9
Resource level
- CulturaX token counts proxy for training exposure, which is undisclosed and may include curated, translated, or synthetic data; resource-level slopes are therefore associations with the proxy.Source: Limitations · p. 9
Data
- Translated, largely Western-centric MMMLU questions may conflate linguistic and cultural distance, while mostly mathematical reasoning tasks may not represent other reasoning forms.Native-language questions would help separate these effects.Source: Limitations · p. 9
Valid generations
- Valid generations fall from 97% in English to 35% in Yoruba, so lower-resource results may reflect a selected subset; robustness checks preserve the knowledge slope but weaken the gap's significance.The authors cannot rule out different behavior among invalid generations.Source: Limitations · p. 9
Measurement
- The metrics test whether answers depend on the CoT, but not whether it reveals everything driving the answer; two interventions also rely on an imperfect perturber.Filler Tokens, which needs no perturber, shows the most robust gap.Source: Limitations · p. 9
Models
- The study evaluates four open-weight models with 4–8B parameters; larger models and reasoning models with long traces may rely differently on reasoning in lower-resource languages.Source: Limitations · p. 9
A Implementation Details
The appendix specifies how the models generate chains and answers, how confidence and faithfulness are measured, and how interventions and robustness controls are applied.
Generation
- Models decode greedily in bfloat16, generate up to 2,048 tokens, and are prompted to reason step by step before placing an answer in a boxed format.CoT prompts are translated into each evaluation language; perturber prompts remain in English.Source: A Implementation Details · p. 12, Table 6 · p. 13
Confidence
- Answer confidence is the teacher-forced probability of the original answer tokens, while chain confidence uses token-level measures across the CoT and answer.The probability measure is omitted for the chain because it vanishes over long sequences.Source: A Implementation Details · p. 12
Interventions
- Four interventions truncate the CoT, replace it with length-preserving filler, introduce an antonym, or paraphrase most of the chain before the model continues.After editing, the original answer is re-appended and scored by teacher forcing.Source: A Implementation Details · p. 12
Paraphrasing as corruption
- Unlike prior work treating Paraphrasing as meaning-preserving, this study treats it as corruption because paraphrases often alter chains and answers.Median similarity is 0.68–0.82 for paraphrases versus at least 0.99 for Adding Mistakes; answer changes occur in 16% versus 8% of cases.Source: A Implementation Details · p. 12
Text similarity
- Chain similarity is computed after Unicode normalization and removal of formatting and whitespace effects; word-level diffs check whether mistake edits are single substitutions.Source: A Implementation Details · p. 13
Robustness controls
- Robustness controls demean faithfulness within groups, including comparing each question with itself across languages for item fixed effects.Other controls group generations into pooled-sample deciles.Source: A Implementation Details · p. 13
B Data Sanity
Generation validity and accuracy decline in Yoruba, while checks of the perturber show that most edits follow instructions but some larger edits warrant controls.
Generations
- 97% of English generations and 35% of Yoruba generations are valid; among valid generations, accuracy falls from 0.67 to 0.31, near chance at 0.25.Mean answer probability falls from 0.99 to 0.93, and near-certain answers from 93% to 63%.Source: B Data Sanity · p. 13, Table 7 · p. 14
Perturber fidelity
- Adding Mistakes edits have median similarity of at least 0.99 in every language, but larger rewrites occur especially in Swahili and Bengali and could inflate faithfulness.The authors control for similarity and check a single-word-edit subset; paraphrases change less in lower-resource languages, which would weaken rather than inflate the effect, apart from a tail of Yoruba rewrites.Source: B Data Sanity · p. 14, Table 8 · p. 14
C Expanded Results
This section extends the main-text analyses across every intervention and metric.
Expanded analyses
- The expanded analyses cover every intervention and metric in the order used in the main text.Source: C Expanded Results · p. 14
C.1 Faithfulness Variation
Faithfulness varies with resource level differently across task families, metrics, and interventions; the knowledge–reasoning gap is generally negative for knowledge-task comparisons.
C.1 Faithfulness Variation
- Across Adding Mistakes and Paraphrasing, knowledge-task faithfulness rises as resource level falls under every metric, with a negative knowledge–reasoning difference in 11 of 12 combinations.Figure 6 repeats the main comparison for ∆p and Φ; Table 9 reports slopes under all three metrics.Source: §C.1 · p. 14, Figure 6 · p. 15, Table 9 · p. 15
- Under Filler Tokens, the knowledge-task slope is significant only for ∆log p; under Paraphrasing, reasoning-task ∆p and Φ also rise as resource level falls, narrowing the family gap.Source: §C.1 · p. 14
- Under Early Answering, reasoning-task slopes are positive for every metric.Source: §C.1 · p. 14
C.2 Metrics and Models
Across these supplementary analyses, metric comparisons, World Religions results, confidence patterns, and robustness checks refine the reported resource-level and task-family findings.
C.2 Metrics and Models
- Across models, ∆p and ∆log p rank generations almost identically (Spearman 0.92–1.00); after answer flips, ∆log p distinguishes interventions while ∆p reaches its ceiling.The figure compares distributions for answers kept versus flipped; Table 10 reports metric correlations.Source: §C.2 · pp. 15–16, Table 10 · p. 15
- Every model pair agrees positively under each metric, and every model has a negative knowledge–reasoning gap except Qwen under Φ (0.01).Source: §C.2 · p. 16, Table 11 · p. 16
C.3 Subtasks and World Religions
- In World Religions, accuracy falls from about 0.8 in English to 0.3 in Yoruba while Adding Mistakes flips rise from 2% to about 10%; this subtask remains steepest.Figure 9 plots accuracy and answer-flip rate against resource level.Source: §C.3 · p. 16, Figure 9 · p. 17, Figure 8 · p. 17
C.4 Confidence Mediation
- In 11 of 13 significant slopes, the expected confidence measure attenuates the resource-level slope more: final-answer probability for knowledge tasks and reasoning-chain entropy for reasoning tasks.The exceptions are Adding Mistakes on knowledge tasks and Paraphrasing on reasoning tasks, both under Φ.Source: §C.4 · p. 17
- Final-answer confidence rises at nearly the same rate in both task families, while reasoning-chain confidence barely changes; this shared gradient cannot explain the faithfulness gap.Final-answer slope differences are at most 0.03 standard deviations.Source: §C.4 · p. 17
C.5 Simpson’s Paradox
- Confidence and faithfulness correlate strongly between languages (mean |r| = 0.47) but nearly vanish within languages (0.06), with |rw| < 0.1 in 164 combinations.The signs differ in 93 combinations, mostly because within-language correlations are near zero rather than clearly opposite.Source: §C.5 · p. 17
C.6 Rate and Magnitude
- The knowledge–reasoning difference appears in rate under Filler Tokens, Adding Mistakes, and Paraphrasing, but in magnitude under Early Answering.Table 15 reports reliance, rate, and magnitude slopes for every intervention.Source: §C.6 · p. 17
C.7 Robustness Tests
Robustness checks show that significance depends more on sample restrictions than on adding controls, with some intervention-specific exceptions.
C.7 Robustness Tests
- Under Filler Tokens, the knowledge–reasoning difference remains significant in 19 of 21 specifications.By contrast, sample restrictions remove significance for the difference under Early Answering and Adding Mistakes.Source: §C.7 · p. 17, Table 16 · p. 21
- Adding a control rarely changes which knowledge–reasoning differences are significant.Source: §C.7 · p. 17
- Table 16 repeats the robustness analysis across interventions and metrics; its perturber checks apply only to interventions using a perturber.Colour is scaled within each metric because the metrics use different units.Source: Table 16 · p. 21