PaperlayerView exact revision on arXiv

Source-linked AI summary

What Do Rationales Communicate? A Message-Intervention Study in Role-Specialized QA

Jiameng Zhang, Hongqiu Wu

arXiv:2610.00018v1cs.AIcs.CL

TL;DR

The study asks what rationale messages change after a candidate answer is proposed: answer selection or support assessment. It varies only the rationale while fixing evidence and candidate answer, finding that corrupted rationales shift support judgments more than faithful rationales improve answer accuracy.

  • Problem

    Once a candidate answer is proposed, it is unclear whether a rationale changes answer selection or the verifier’s judgment that the answer is supported.

    Source: §1 · p. 1

  • Method

    The message-intervention protocol fixes the evidence and candidate answer while varying the rationale sent from reasoner to verifier.

    Source: §1 · p. 1, §2.1 · p. 2

  • Results

    Across three datasets, harmless paraphrases changed support in 0–2.5% of examples, while corrupted rationales changed it in 10–22% under a blind prompt and 34–55% with rationale checking.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Rationales act more strongly as verification messages than as answer-selection signals, so their effects on support judgments warrant evaluation beyond answer accuracy.

    Source: §1 · p. 2

    Most conditions assume a candidate answer is formed before verification; the answer-free stress test and split-channel check are limited to HotpotQA-100.

    Source: §7 · p. 9

Abstract

from arXiv · show

Role-specialized QA pipelines increasingly pass rationales from a reasoner to a verifier, but it is unclear what this message actually buys: better answers, stronger support assessment, or a new failure surface. We introduce a message-intervention diagnostic that fixes the evidence and candidate answer while varying only the rationale passed across the reasoner-to-verifier boundary. On 400 MuSiQue, HotpotQA, and 2WikiMultiHopQA examples with DeepSeek as generator and verifier, faithful rationales add almost no answer accuracy over no rationale, while corrupted rationales strongly alter support judgments. Under a blind verifier prompt, harmless paraphrases shift support by only 0--2.5%, whereas corrupted rationales shift support by 10--22%; an explicit rationale-checking prompt amplifies the same pattern to 34--55%. Final answers move less (2--30%), and only 2.9--35.3% of corrupted support flips co-occur with answer changes. Human audits show why this matters: 16/42 valid corruptions are corruption-overtrust cases, and blind humans reject or mark unclear 9/10 audited corrupted rationales that the model accepts. Cross-model and task-boundary checks show when the channel is active, amplified, inert, or folded into the task label. Rationale sharing should be evaluated as a verification-message mechanism, not merely as a route to higher answer accuracy.

1 Introduction

Role-specialized QA pipelines pass a reasoner’s rationale to a verifier, but end-task accuracy cannot reveal whether it changes answer selection or support assessment. The paper reframes this as an intervention on the rationale message.

1 Introduction

  • The unresolved interface question is whether a rationale improves answer selection or mainly changes support assessment, which end-task accuracy cannot distinguish.A verifier returns both an answer and a judgment that the answer is supported.Source: §1 · p. 1
  • The intervention fixes evidence and candidate answers, varies the rationale passed to the verifier, and measures answer selection separately from support assessment.Tested rationale conditions include no rationale, the original rationale, a meaning-preserving paraphrase, and two corruption types.Source: §1 · pp. 1–2
  • Faithful rationales add only +0.5 EM on MuSiQue over passing evidence and answer alone, while corrupted rationales change support 10–22% blind and 34–55% with faithfulness checks.The diagnostic contrasts answer accuracy with support-assessment sensitivity across three multi-hop QA datasets.Source: §1 · p. 2
  • Blind humans reject or mark unclear 9/10 audited corrupted-supported cases; the paper also identifies correct-answer penalty and local answer dominance.Local answer dominance occurs when direct evidence preserves an answer despite a broken upstream rationale.Source: §1 · p. 2
  • The diagnostic turns rationale sharing into a measurable question and uses human audits and cross-model, prompt, noise-floor, and task-extension checks to characterize the channel.The paper asks what behavior the rationale message changes.Source: §1 · p. 2

2 Diagnostic Framework

The framework treats a rationale as an explicit message between specialized QA roles and isolates its effect by changing only that message. It measures answer and support behavior separately across wording-preserving and factual-corruption conditions.

2.1 Role-Specialized QA Pipeline

  • The retriever selects evidence, the reasoner proposes an answer and rationale, and the verifier returns a final answer plus a binary support judgment.The intervention holds the task, evidence, candidate answer, verifier role, and output schema fixed, varying only the rationale field.Source: §2.1 · p. 2

2.2 Perturbation Conditions

  • Harmless paraphrases preserve entities, relations, dates, and answer support, while corruptions swap key facts or imply a different plausible answer.These conditions distinguish wording brittleness from responsiveness to factual errors.Source: §2.2 · pp. 2–3

2.3 Metrics and Testing

  • The metrics track answer accuracy, support rate, paired answer and support sensitivities, and how often answer changes accompany support flips.Harmless robustness means low support sensitivity under paraphrase; corruption responsiveness means high support sensitivity under factual corruption.Source: §2.3 · p. 3
  • Paired McNemar tests compare original-rationale support judgments with harmless and corrupted conditions.Source: §2.3 · p. 3

3 Experimental Setup

The study evaluates three multi-hop QA datasets using a DeepSeek reasoner and verifier, with SciFact as a task-boundary check. Human audits assess intervention quality and whether rationale errors also occur in unperturbed outputs.

3.1 Datasets and Models

  • The main evaluation covers 400 examples across MuSiQue, HotpotQA, and 2WikiMultiHopQA, with a smaller SciFact claim-verification extension.The main pipeline uses DeepSeek as generator/reasoner and verifier; the extension tests a task whose output is itself a support/refute label.Source: §3.1 · p. 3
  • Verifier-note analysis uses deterministic string checks for mentions of the corrupted target and broader conflict terms.The target-aware check is stricter evidence that the verifier localizes the introduced semantic conflict.Source: §3.1 · p. 3

3.2 Perturbation and Human Audits

  • Audits found 42/48 corrupted rationales valid, 36/50 paraphrases strictly valid, and 43/50 acceptable under a lenient criterion.A blind human support audit supplements the perturbation checks; a separate audit found 11 unfaithful rationales among 100 unperturbed outputs.Source: §3.2 · pp. 3–4

4 Rationales Verify More Than They Answer

Across the QA results, factual corruption changes support judgments much more than harmless rewording, while final answers are often more stable. The analyses distinguish these outputs, test how prompt and task structure affect the pattern, and explain why a rationale can alter support without changing an answer.

4.1 Harmless and No-Rationale Baselines

  • No-rationale and faithful-rationale conditions were close: EM differed by at most one point and support rates by at most three points.Harmless changes were not significant, with the HotpotQA faithful support rate already at 100%.Source: §4.1 · p. 4

4.3 Semantic Corruption Changes Support Judgments

  • 34–55% of examples changed support under entity-swapped rationales, versus 35–51% under answer-conflicting rationales; harmless paraphrases changed support in only 4%, 1%, and 2% across the datasets.Corrupted-condition comparisons were significant on every dataset; support rates fell from 85.5–100.0% under faithful rationales to 42.0–66.0% under corrupted rationales.Source: §4.3 · p. 4, §4.1 · p. 4, Table 1 · p. 5

4.4 Final Answers Stay More Stable Than Support

  • Only 2.9–35.3% of corrupted-rationale support flips accompanied answer changes, and answer-free HotpotQA checks still showed larger support than answer movement.In the answer-free arm, entity swaps moved support in 60% versus answers in 13%, while answer conflicts moved support in 50% versus answers in 22%.Source: §4.4 · pp. 4–5, Table 1 · p. 5

4.2 Blind Prompt: The Effect Is Not Mere Prompt Compliance

  • Under a blind prompt, harmless paraphrases changed support in 0–2.5% of examples, while corrupted rationales changed it in 10–22%.The effect is weaker than under explicit rationale-faithfulness instructions, but remains across all three QA datasets.Source: §4.2 · p. 4, Table 2 · p. 5

4.5 The Support Bit Combines Answer and Rationale Checks

  • On HotpotQA-100, corrupted rationales left answer-support judgments at 90% and 84%, while rationale-faithfulness judgments fell to 12% and 11%.Overall support followed rationale faithfulness at 12% and 9%, suggesting that the single support bit can combine answer support and rationale faithfulness.Source: §4.5 · pp. 5–6

4.6 Why the Dissociation Happens

  • Directly recoverable answer spans can remain unchanged despite rationale corruption, while a verifier assessing the full output may reject a rationale that contradicts evidence.The diagnostic distinguishes appropriate bundle rejection from correct-answer penalty and corruption overtrust, which final EM alone cannot reveal.Source: §4.6 · p. 6

4.7 Depth and Task Structure Bound the Claim

  • Support sensitivity rose with hop count in MuSiQue, while SciFact’s support/refute output showed label changes of 34% and 45% under the two corruption types.The passage frames the task boundary as a difference between extractive answer selection and a final label that is itself a support judgment.Source: §4.7 · p. 6
  • The design implication is to test whether a rationale channel is active, harmless, or ignored, then distinguish rationale repair, evidence rechecking, support escalation, and answer regeneration.When the channel is ignored, passing rationales may add interface complexity without communication value.Source: §4.7 · pp. 6–7

5 Validation and Robustness

Across audits and cross-model checks, corrupted rationales can shift verifier support judgments even when harmless paraphrases do not. DeepSeek-R1 is a boundary case: it largely discounts the rationale channel.

5.1 Audits and Cross-Model Checks

  • 16/42 valid corruptions were corruption-overtrust cases, while 13/42 were correct-answer penalty cases; blind humans rejected or marked unclear 9/10 model-accepted corruptions.The remaining 13/42 cases involved changed answer correctness or appropriate rejection.Source: §5.1 · p. 7
  • 0% harmless support change versus 39.5% for entity swaps and 34.9% for answer conflicts after acceptable-paraphrase filtering; both corrupted conditions changed support in 36.1% of strict-valid cases.The full audited sample had 2% harmless support change.Source: §5.1 · p. 7
  • 1.5% of faithful-rationale support bits changed on retest, compared with 34–55% corrupted-rationale changes, while EM changed in 0/400 examples.The retest used the same model, prompt, inputs, and temperature.Source: §5.1 · p. 7
  • Qwen-verifier checks showed 0% harmless versus 75–77% corrupted support change; Claude showed 2% versus 28%, and Qwen-generated rationales with DeepSeek verification showed 0% versus 32–34%.The pattern generalized across the tested verifier and generator substitutions, with different effect sizes.Source: §5.1 · p. 7

5.2 Reasoning Verifiers Can Make the Channel Inert

  • DeepSeek-R1 had 96% answer stability and only 2–8% corrupted-rationale support change, indicating channel inertia rather than strong rationale influence.It showed no harmless support changes and may discount the external rationale field.Source: §5.2 · p. 7

6 Related Work

Prior work studies explanation faithfulness, verifier-based selection, judge sensitivity, and communication among agents. This paper focuses on a single rationale message crossing a reasoner–verifier boundary and separates answer selection from support assessment.

6 Related Work

  • Faithfulness research tests whether rationales reflect a model’s decision process, while this work examines a rationale after it crosses a role boundary and separates answer selection from support assessment.The related work also notes that some perturbation tests measure output-level self-consistency rather than hidden reasoning.Source: §6 · p. 7
  • Verifier-based reasoning uses verifiers, self-consistency, and process supervision to select or check solutions, while judge and QA reevaluation studies examine sensitivity to prompts, references, answer forms, and rubrics.RAG evaluation separately considers faithfulness, answer relevance, and context use.Source: §6 · p. 8
  • Multi-agent communication studies information exchange, error cascades, and harmful agreement across agents or debate rounds; this paper narrows the interface to one rationale field between a reasoner and verifier.Message perturbations test whether the receiver treats that field as answer evidence, support evidence, or inert text.Source: §6 · p. 8

7 Limitations

The diagnostic isolates rationale-message effects under fixed evidence and candidate answers, so its conclusions concern interface behavior rather than jointly changing pipeline stages. Its scope is further bounded by study scale, task coverage, sampled audits, and black-box behavioral evidence.

7 Limitations

  • Holding evidence and candidate answer fixed isolates message effects, while end-to-end evaluations remain complementary when retrieval, reasoning, answer generation, and verification change together.The verifier’s binary support label is a prompt- and evidence-context-specific decision, not a formal proof label.Source: §7 · pp. 8–9
  • 400 primary examples across three datasets prioritize controlled comparisons and statistical rigor over benchmark scale; the paired intervention also requires multiple verifier calls and perturbation variants per example.Perturbations are LLM-generated and audited on samples rather than exhaustively labeled.Source: §7 · p. 8
  • The primary experiments study multi-hop QA, with SciFact as a smaller extension, so conclusions are strongest for role-specialized QA and verification pipelines.Tool use, planning, and open-ended collaborative agents are proposed as future targets for the diagnostic.Source: §7 · pp. 8–9
  • Explicit rationale-checking prompts amplify the channel, while blind prompts show whether it remains active without that instruction.The main prompt models systems that intentionally pass rationales for verification.Source: §7 · p. 9
  • These are black-box behavioral tests; explaining the internal computations behind response changes, especially for DeepSeek-R1, remains a separate mechanistic question.Possible explanations include independently reconstructing support, different prompt priors, or ignoring external rationale fields.Source: §7 · p. 9
  • Most conditions assume the candidate answer is formed before verification; the answer-free stress test and split-channel check are limited to HotpotQA-100.The split-channel check is a mechanism check and template for future multi-dataset probes.Source: §7 · p. 9

A.1 Human Support Audit

The human support audit uses a blind annotator on an enriched set of diagnostic cases, rather than a sample designed to estimate global verifier accuracy.

A.1 Human Support Audit

  • The 50-item audit contains 20 corrupted cases where the model rejects support, 10 corrupted cases where it still accepts support, 10 harmless controls, and 10 faithful baselines.The blind annotator saw the question, evidence, answer, and displayed rationale, but not model labels or condition metadata.Source: §A.1 · p. 11
  • Blind model-human agreement was 74% with κ = 0.52 across three labels; disagreement concentrated in corrupted-supported cases the human rejected.The two human passes agreed on 88% of items, or 93.6% after excluding unclear labels.Source: §A.1 · p. 11
  • Table 3 reports blind human support audit results by enriched sampling bucket.The buckets are diagnostic cases, harmless controls, and faithful baselines rather than a global-accuracy sample.Source: Table 3 · p. 11, §A.1 · p. 11

A.2 Harmless-Paraphrase Filtering

Audit-filtered harmless paraphrases leave support judgments unchanged, while corrupted rationales remain disruptive. A separate audit finds rationale inconsistencies in unperturbed reasoner outputs, alongside evaluation artifacts and correct rejections.

A.2 Harmless-Paraphrase Filtering

  • 0% of original-to-harmless support changes remain in both acceptable and strict-valid subsets after filtering, while corrupted rationales remain disruptive.Table 4 summarizes audit-filtered support-change rates relative to the original rationale condition.Source: §A.2 · p. 11, Table 4

A.3 Natural Rationale Inconsistency Audit

  • In 100 unperturbed outputs, 66 were consistent and 11 contained unfaithful rationale errors.The audit sampled rationales across MuSiQue, HotpotQA, and 2Wiki.Source: §A.3 · p. 11, Table 5 · p. 11
  • Among the remaining audit labels, 16 were metric false negatives, 5 correct rejections of wrong answers, and two other cases were verifier or question/gold flaws.The authors say the audit shows the intervention studies a naturally occurring failure mode while guarding against attributing ordinary EM errors to rationale unfaithfulness.Source: §A.3 · p. 11

B Qualitative Case Details

The case examples distinguish answer rejection, corruption overtrust, and local answer dominance. Stress tests show that support remains more sensitive than answers without a supplied answer, while a split-channel probe localizes corruption effects mainly to rationale faithfulness and overall support.

B Qualitative Case Details

  • The examples contrast a correct answer rejected after a corrupted rationale, a corrupted link whose answer remains supported, and a locally supported answer spared from an upstream relation error.Table 6 presents these as two auditable failure modes and one qualitative dissociation pattern.Source: B Qualitative Case Details · p. 11, Table 6 · p. 11

C Answer-Free Stress Test

  • Without a supplied candidate answer, corrupted rationales still move support more than answers under both corruption conditions.Table 7 reports the HotpotQA-100 stress test; removing the answer anchor increases answer sensitivity, especially for answer-conflicting rationales.Source: C Answer-Free Stress Test · p. 12, Table 7 · p. 12

D Split-Channel Verification Details

  • In the split-channel probe, corrupted rationales mostly affect rationale-faithfulness and overall-support judgments, while answer support remains substantially higher.Table 8 reports the HotpotQA-100 probe that separately elicits answer support, rationale faithfulness, and overall support.Source: D Split-Channel Verification Details · p. 12, Table 8 · p. 12

E Additional Visualizations

The additional visualizations make the answer/support stability gap visible, show that corruption reduces support while harmless paraphrases track faithful rationales, and include a task boundary where support sensitivity reaches the final label. The accompanying details specify hosted-model settings and the fixed-message intervention design.

E Additional Visualizations

  • Multi-hop QA shows support sensitivity with relatively stable final answers, whereas SciFact is a boundary case where support changes propagate into the final label.Figure 3 provides the cross-task support-change view.Source: Figure 3 · p. 13, E Additional Visualizations · p. 13
  • Harmless paraphrases closely track faithful rationales, while both corruption types reduce verifier support rates.Figure 4 visualizes raw supported rates by rationale condition.Source: E Additional Visualizations · p. 13, Figure 4 · p. 13
  • Answers remain more stable than support judgments across the five rationale conditions, with the strongest support instability on MuSiQue and HotpotQA.Figure 5 shows the fraction of examples whose answer and support judgment remain unchanged across conditions.Source: E Additional Visualizations · p. 13, Figure 5 · p. 13
  • After filtering to valid or strict-valid paraphrases, harmless changes disappear while corrupted-rationale changes remain large.Figure 6 visualizes the audit-filtered harmless-control comparison.Source: E Additional Visualizations · p. 13, Figure 6 · p. 13

F Model and Decoding Details

  • The experiments report hosted API model identifiers and temperature-0 decoding, but provider-side parameter counts and exact weight snapshots are undisclosed.Table 9 lists model identifiers and decoding settings.Source: F Model and Decoding Details · p. 13, Table 9 · p. 14

G Verifier Prompt and Output Schema

  • The intervention holds evidence, candidate answer, verifier, decoding, and output schema fixed while varying only the rationale message; a blind-prompt ablation removes explicit rationale-faithfulness instructions.The verifier returns a final answer, binary support label, confidence, and short verification note.Source: G Verifier Prompt and Output Schema · p. 14

Related papers

Loading 2610.00018v1…