PaperlayerView exact revision on arXiv

Source-linked AI summary

Right Answers, Wrong States: Hidden Information Failures in Multi-Agent Collaboration

Herun Wan, Jiaying Wu, Minnan Luo, Zihan Ma, Fanxiao Li, Nancy F. Chen, Min-Yen Kan

arXiv:2610.01244v1cs.CL

TL;DR

Correct current answers can leave corrupted information in a shared state, a problem that task accuracy alone may miss. The paper introduces OFFQUERY to measure evidence verification, state reconstruction, and task resolution, and REGROUND to verify evidence, repair shared facts, and reason over the reconstructed state. REGROUND improves all three capabilities across evaluated model–setting combinations, while longer-horizon persistent-memory dynamics remain outside the benchmark’s scope.

  • Problem

    Multi-agent systems can answer the current query correctly while leaving unresolved errors in shared information that later reasoning may reuse.

    Source: §1 · p. 1

  • Method

    OFFQUERY separates evidence verification, shared-state reconstruction, and task resolution, while REGROUND verifies conflicting evidence, reconstructs trusted shared information, and reasons over that state.

    Source: §1 · p. 2, §4 · p. 5

  • Results

    REGROUND improves all three tasks in every model–setting combination, with average relative gains of 309.0% on T1, 82.9% on T2, and 17.6% on T3.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings support evaluating the reliability of information carried forward, in addition to whether the current answer is correct.

    Source: §7 · p. 9

    OFFQUERY evaluates a shared state within a bounded collaboration episode, not persistent memory and state updates across longer-horizon interactions.

    Source: A LIMITATIONS AND FUTURE WORK · p. 14

Abstract

from arXiv · show

Multi-agent systems are often judged by whether they reach the correct answer. This can miss a distinct failure: collaboration may leave behind a corrupted information state even when the immediate decision is correct. We call this an off-query failure. To study this failure in collaborative decision support, we introduce OffQuery, which separately evaluates evidence verification (T1), shared-state reconstruction (T2), and task resolution (T3) in two representative high-stakes settings: healthcare and disaster response. Across GPT, Gemini, and Qwen models, standard collaboration shows much stronger task performance than state reliability. Averaged over 21 model--setting combinations, task resolution reaches 64.7%, while evidence verification and state reconstruction reach only 14.3% and 43.1%. We trace this gap to selective information use: current queries often bypass corrupted facts, which become consequential when later tasks require them. We further introduce ReGround, which resolves conflicting evidence, verifies shared facts, reconstructs a trusted state, and reasons over that state. Across seven models from three families, ReGround improves all three capabilities in every evaluated setting, with average relative gains of 309.0%, 82.9%, and 17.6% on T1, T2, and T3. Reliable collaboration therefore requires both a correct decision and a reliable shared state for future reasoning.

1 INTRODUCTION

Correct answers can conceal corrupted shared information when the current query does not use it. OFFQUERY measures this gap, while REGROUND explicitly verifies evidence and reconstructs state to improve reliability and task performance.

1 INTRODUCTION

  • Off-query failures occur when a correct current answer leaves errors in the shared state that remain hidden until later reasoning reuses them.Task accuracy audits only the portion of the state exercised by the current query.Source: §1 · p. 1
  • Across 21 model–setting combinations, standard collaboration averages 64.7% on T3, versus 14.3% on T1 and 43.1% on T2.The gap persists across GPT, Gemini, and Qwen models, scales, and both high-stakes domains.Source: §1 · p. 2
  • Selective information use helps explain the gap: corrupted facts appear less often in explanations for correct decisions, while 43% of broader clinical plans from corrupted states conflict with plans from the reliable reference state.Direct corruption lowers downstream accuracy, while progressively repairing the state improves performance.Source: §1 · p. 2
  • REGROUND improves T1, T2, and T3 across all 21 model–setting combinations, with average relative gains of 309.0%, 82.9%, and 17.6%, respectively.Its largest gains are in evidence verification and state reconstruction, where standard collaboration is weakest.Source: §1 · p. 2

2 RELATED WORK

Prior multi-agent work studies collaboration strategies and their effects on task outcomes, while other studies document information failures. OFFQUERY focuses on whether errors remain in the shared state after a current query succeeds.

Multi-Agent Collaboration

  • Multi-agent systems combine complementary expertise through role specialization, structured interaction, and iterative debate (Feng et al., 2026; Li et al., 2023; Chen et al., 2024b; 2025; Hong et al., 2024; Wu et al., 2024; Chen et al., 2024a; Du et al., 2024).Work also examines how interaction structure, participation, and context management shape collaborative performance.Source: §2 · pp. 2–3

Information Failures in Multi-Agent Collaboration

  • Prior studies report that debate can converge on misconceptions, adversarial or sycophantic agents can distort decisions, and misleading information can propagate through interaction (Li et al., 2026a; Yan et al., 2026).OFFQUERY studies the distinct case where the current task is answered correctly while errors remain in the shared state.Source: §2 · p. 3

Evaluating Collaborative State

  • Existing benchmarks assess task success, coordination, information integration, or fact survival (Wan et al., 2026).OFFQUERY separately tests evidence verification, shared-state reconstruction, and task resolution against a known global information state.Source: §2 · p. 3

3 OFFQUERY: EVALUATING INFORMATION CORRECTNESS OF MULTI-AGENT COLLABORATION

OFFQUERY evaluates whether multi-agent collaboration can identify unreliable evidence, reconstruct a reliable shared state, and answer the current query. Its healthcare and disaster-response benchmark uses controlled information distribution and corruption, with validation of construction and evaluation reliability.

3.1 TASK FORMULATION

  • T1 evaluates identification of unreliable evidence, T2 evaluates reconstruction of a reliable shared state, and T3 evaluates the current query; errors in earlier stages can affect later ones.High T3 can coexist with poor T2 when the current query accesses only part of the shared state.Source: §3.1 · pp. 3–4

3.2 BENCHMARK CONSTRUCTION

  • OFFQUERY uses MediQ and CrisisFacts to build healthcare and disaster-response instances with a known reliable global state and distributed information.Clinical information is partitioned by category; disaster-response facts are assigned to incident-relevant stakeholders, with partitions jointly covering the reliable context.Source: §3.2 · p. 4
  • The benchmark injects plausible misleading evidence into private or shared information, designed to conflict with the reliable state and affect downstream reasoning if accepted.In the main setting, misleading private evidence is assigned to one agent.Source: §3.2 · p. 4

3.3 EVALUATION AND QUALITY VALIDATION

  • T1 uses source-identification accuracy, T2 checks preservation of reliable atomic facts and removal of misleading ones, and T3 uses multiple-choice task accuracy.The T2 evaluation reduces sensitivity to differences in wording, ordering, or level of detail.Source: §3.3 · p. 4
  • With misleading private information, GPT-4o accuracy falls to 12.6% and 6.0% on HEALTHCARE-S and HEALTHCARE-I; with misleading shared information, it falls to 20.2% and 5.3%, respectively.Manual inspection of 50 sampled instances found no factual omissions or hallucinations relative to the source context.Source: §3.3 · pp. 4–5
  • T2 judge-model aggregate scores have Spearman rank correlation 1.00, pairwise instance-level mean absolute errors of 0.126–0.139, and average human–GPT-4o agreement of 0.86 (Cohen’s κ=0.84).The results indicate stable evaluation across judge models and close alignment with human assessment.Source: §3.3 · p. 5

4 REGROUND: RELIABLE SHARED-STATE RECONSTRUCTION

REGROUND addresses unreliable shared information through evidence verification, atomic-fact state reconstruction, and reasoning over the reconstructed state. Its design resolves conflicts before facts enter the trusted context used for downstream decisions.

4 REGROUND: RELIABLE SHARED-STATE RECONSTRUCTION

  • REGROUND aligns three stages with OFFQUERY’s tasks: verify evidence, reconstruct the shared state, and resolve the task using that reconstructed state.The design principle is to resolve conflicting information before it becomes trusted shared context.Source: §4 · p. 5

4.1 CONFLICT-GUIDED EVIDENCE VERIFICATION

  • REGROUND withholds the potentially unreliable shared context during verification, then discovers candidate sources through iterative conflict resolution and scores candidates against evidence from outside the candidate pool.This separates finding where verification is needed from deciding which source to trust.Source: §4.1 · p. 5

4.2 EVIDENCE-GROUNDED STATE RECONSTRUCTION

  • After verification, REGROUND decomposes the initial shared context into atomic facts, retains supported facts, removes contradicted or unsupported ones, and consolidates the survivors with trusted private evidence.The resulting state records information supported by verified evidence rather than carrying forward claims irrelevant to the current query.Source: §4.2 · p. 6

4.3 REASONING OVER THE RECONSTRUCTED STATE

  • REGROUND answers query q from reconstructed state ˆC, conditioning all task reasoning on the state established by verification and reconstruction.This design makes downstream decisions use the preceding stages’ shared-state output.Source: §4.3 · p. 6

5 EXPERIMENTS

Across the experiments, standard collaboration often answers correctly while leaving an unreliable shared state; ReGround improves evidence verification, state reconstruction, and task resolution. The results also show that distributed information and unreliable shared context particularly burden state reconstruction.

5.1 EXPERIMENTAL SETUP

  • Experiments compare standard collaboration and ReGround with an Oracle across HEALTHCARE-S, HEALTHCARE-I, and multiple GPT, Gemini, and Qwen model scales.The study also compares ReGround with voting, sequential, centralized, self-reflection, and confidence-based strategies.Source: §5.1 · p. 6

5.2 MAIN RESULTS

  • 86.2% on T3 but only 10.9% on T1 and 33.2% on T2 for GPT-5 on HEALTHCARE-S exemplifies how task success can overstate state reliability.The separation persists across model families, scales, and evaluation settings.Source: §5.2 · p. 6, Table 1 · p. 7
  • 34.2 and 34.3% are the T2 drops from Oracle with reliable distributed information, versus T3 drops of 6.9 and 12.4%.With unreliable shared context, T2 degrades by 65.9 and 66.6% versus Oracle, compared with 36.9 and 38.1% when unreliability is confined to private evidence.Source: §5.2 · p. 6, Table 2 · p. 7
  • 309.0% (T1), 82.9% (T2), and 17.6% (T3) are ReGround’s average relative gains over standard collaboration across 21 model–setting combinations.Improvements occurred across every model family, scale, and evaluation setting.Source: Table 1 · p. 7, §5.2 · p. 7
  • Explicit evidence verification and state maintenance can address part of the collaboration failure without relying only on stronger task-solving models.ReGround determines what information should survive collaboration before it is reused for reasoning.Source: §5.2 · p. 7
  • ReGround achieves the highest T3 on all three datasets across GPT-4o and Gemini-3.5-Flash, complementing strategies that change agent interaction or aggregation.The comparison covers voting, sequential interaction, centralized coordination, self-reflection, and confidence-based aggregation.Source: §5.2 · p. 8

6 WHY OFF-QUERY FAILURES OCCUR AND WHEN THEY MATTER

Narrow queries can bypass corrupted parts of a shared state, leaving errors latent until later tasks reuse that information. Experiments show that broader reuse exposes such errors, while delaying exposure to potentially corrupted shared context and explicitly maintaining state can improve reliability.

6 WHY OFF-QUERY FAILURES OCCUR AND WHEN THEY MATTER

  • ReGround improves downstream task resolution across both model backbones and all three settings in Table 3.D1–D3 denote HEALTHCARE-S, HEALTHCARE-I, and DISASTER RESPONSE.Source: Table 3 · p. 8
  • 43% of broader clinical plans were inconsistent with reliable-reference plans despite all 100 instances succeeding on the original narrow task.The broader assessment covered diagnosis, alternatives, immediate management, and next steps.Source: §6 · p. 8
  • Corrupting private or shared information substantially reduces downstream T3 accuracy, while progressively more reliable states generally yield higher accuracy.The state-quality interventions remove inter-agent communication as an explanation; accuracy reaches its highest or near-highest level with fully reliable states.Source: §6 · pp. 8–9, Figure 3 · p. 9
  • ReGround improves performance from 45.2% to 60.8% with multiple misleading agents and from 26.4% to 49.7% with misinformation distributed across agents.These harder settings reduce the usefulness of identifying a single bad source, while explicit verification and reconstruction continue to improve performance.Source: §6 · p. 9
  • Introducing shared context before discussion offers little gain when reliable but consistently degrades performance when corrupted.The alternative is to discuss private evidence first, then introduce shared context.Source: Figure 4 · p. 9, §6 · p. 9

7 CONCLUSION

The paper introduces OFFQUERY to assess hidden information failures and REGROUND to explicitly verify evidence and reconstruct shared state. Its findings show a persistent gap between task success and state reliability, while the study is scoped to research evaluation rather than clinical or emergency-response advice.

7 CONCLUSION

  • OFFQUERY evaluates hidden information failures, while REGROUND explicitly verifies evidence and reconstructs shared state.Across seven models and two high-stakes domains, OFFQUERY reveals that systems can answer correctly while leaving errors unresolved in the shared state.Source: §7 · p. 9
  • REGROUND consistently improves evidence verification, state reconstruction, and downstream task performance across the evaluated settings.The conclusion calls for treating information carried forward after collaboration as a reliability target alongside the current answer.Source: §7 · p. 9
  • The study uses existing healthcare and disaster-response datasets with synthetically generated unreliable information, and its systems are not clinical or emergency-response advice.The benchmark and resulting systems are intended for research on model reliability.Source: ETHICS STATEMENT · p. 10

A LIMITATIONS AND FUTURE WORK

This block describes OFFQUERY’s scope and construction: it adapts healthcare and disaster-response data into distributed, unreliable-information tasks. It also identifies limits to the benchmark’s domains, corruption assumptions, and bounded collaboration episodes.

A LIMITATIONS AND FUTURE WORK

  • OFFQUERY covers healthcare and disaster response, uses controlled corruption that can include one misleading private source, and evaluates states within a bounded collaboration episode.Broader domains with ambiguous or evolving ground truth, claim-level reliability without source identifiability, and persistent memory across tasks remain directions for study.Source: A LIMITATIONS AND FUTURE WORK · p. 14

B.1 SOURCE BENCHMARKS

  • MediQ retains 1,310 long, detailed patient cases, while CrisisFact converts 18 disasters into event-day examples and yields 93 examples after preprocessing.MediQ examples have contexts over 815 characters and 10 sentences; CrisisFact filters facts with importance scores below 0.95.Source: §B.1 · pp. 14–15
  • For CrisisFact, each event-day becomes a multiple-choice response question with one optimal strategy and three plausible hard negatives that depend on hidden facts.The hard negatives are designed to be reasonable with partial information but suboptimal with the complete context.Source: §B.1 · p. 15

B.2 BENCHMARK GENERATION

  • MediQ partitions cases into nine departments plus “other,” sharing demographics and chief complaints; CrisisFact uses adaptive stakeholder-based partitions.The CrisisFact prompt calls for a shared situation overview and three to five additional sections, with facts assigned once across sections.Source: §B.2 · pp. 15–16
  • OFFQUERY generates misleading private information and incorrect decisions, then uses incorrect decisions to construct shared contexts that support an alternative explanation.Fabricated private information is intended to be plausible but inconsistent with, and refutable by, the complete context; shared contexts add necessary information without stating the answer or strategy.Source: §B.2 · pp. 16–17

B.3 DATASET STATISTICS

  • Healthcare is divided into HEALTHCARE-S (723 instances) and HEALTHCARE-I (587 instances), according to whether GPT-4o answers from the original shared patient information.The split criterion is whether GPT-4o can answer correctly using only that original shared information.Source: §B.3 · p. 17, Table 4 · p. 17

B.4 BENCHMARK EXAMPLE

  • Figure 5 combines partial reliable shared context, plausible unreliable shared context, complementary private observations, and one misleading observation aligned with the incorrect answer.The instance requires cross-agent evidence aggregation while remaining robust to conflicting information.Source: §B.4 · p. 17, Figure 5 · p. 18

B.5 METRIC DEFINITION

OFFQUERY defines three complementary metrics to separate source verification, shared-state reliability, and current-query accuracy.

T1: Evidence Verification

  • T1 identifies the single misleading agent per instance, using exact-match accuracy to measure whether the system attributes unreliable information to its source.Each instance contains exactly one agent providing misleading information.Source: §B.5 · p. 18

T2: State Reconstruction

  • T2 measures reconstructed-state reliability over atomic facts: positive facts should be retained and misleading negative facts excluded, regardless of wording or order.Its score averages per-instance accuracy on binary labels for whether facts are present or supported.Source: §B.5 · pp. 18, 20
  • When atomic-fact counts exceed task-specific limits, evaluation retains at most 10 facts for HEALTHCARE and 16 for DISASTER RESPONSE instances.GPT-4o is prompted to generate the atomic facts.Source: §B.5 · p. 19

T3: Task Resolution

  • T3 measures conventional downstream decision accuracy by comparing each system prediction with the single correct answer for a multiple-choice problem.T1 and T3 use exact-match accuracy.Source: §B.5 · pp. 18, 20

Metric overview

  • Together, T1 asks whom not to trust, T2 what the group should believe, and T3 whether it makes the correct decision.This decomposition distinguishes correct answers from reliable collective information states.Source: §B.5 · p. 20

B.6 LLM-AS-A-JUDGE CONSISTENCY

The three LLM judges agree on the ranking of evaluation settings and show small instance-level score differences. Human validation also finds GPT-4o judgments largely consistent with human evaluation.

B.6 LLM-AS-A-JUDGE CONSISTENCY

  • Table 6 shows small pairwise MAEs across judge pairs and settings, with average values of 0.142, 0.133, and 0.127 on HEALTHCARE S and 0.135, 0.131, and 0.125 on HEALTHCARE I.The pairs are GPT-4o vs. Gemini-3.5-Flash, GPT-4o vs. Qwen3-235B-A22B, and Gemini-3.5-Flash vs. Qwen3-235B-A22B, respectively.Source: Table 6 · p. 21, §B.6 · p. 21
  • All three judge pairs yield ρagg = 1.00, producing the same ranking of the seven evaluation settings despite differences in absolute scoring scales.The shared ordering places Oracle first, followed by REGROUND with Gemini-3.5-Flash, GPT-4o, and Qwen3-32B.Source: §B.6 · p. 21

B.7 HUMAN VALIDATION

  • Human validation uses 200 context–fact pairs from 50 sampled instances, with fact polarity and the balanced label composition hidden from annotators.Each instance contributes four facts: two positive and two negative.Source: §B.7 · p. 22
  • GPT-4o agrees with the two human annotators at 0.88 and 0.84, while the human annotators agree at 0.92.The corresponding Cohen’s κ values are 0.76, 0.68, and 0.84; the authors conclude GPT-4o judgments are largely consistent with human evaluation.Source: Table 7 · p. 22, §B.7 · p. 22

C REGROUND IMPLEMENTATION DETAILS

This section specifies ReGround’s prompts, evaluation setups, collaboration procedures, and baseline implementations, including scenario-specific adaptations for healthcare and disaster response. It also describes controlled information-reliability settings that isolate unreliable private or shared evidence while distributing task-relevant information across agents, and reports strategy performance across those settings.

C REGROUND IMPLEMENTATION DETAILS

  • For task resolution, ReGround prompts an LLM to choose an option directly from the reconstructed patient or disaster-situation context.The prompts retain consistent objectives and formats across healthcare and disaster response, with minor scenario-specific adaptations.Source: C REGROUND IMPLEMENTATION DETAILS · pp. 22, 28, 29
  • ReGround extracts factual statements, identifies the most contradictory source pair, gathers supporting and refuting evidence, and adjudicates which source is misleading.It repeats this process for M agents, removing each selected agent from later iterations; the main experiment sets M to 3.Source: C REGROUND IMPLEMENTATION DETAILS · pp. 22–25
  • After selecting M candidate agents, ReGround scores their statements against facts from non-candidates, identifies the likeliest misleading agent, and adds the others’ statements to the reliable evidence set.The final adjudicator considers consistency scores and explanations as evidence, not absolute truth.Source: C REGROUND IMPLEMENTATION DETAILS · pp. 25–27
  • For state reconstruction, ReGround decomposes shared information into atomic facts, verifies each against the evidence set, and summarizes supported facts into a natural-language information state.The verification prompt also assesses each fact’s clinical or operational importance.Source: C REGROUND IMPLEMENTATION DETAILS · pp. 27–28

D.1 HEALTHCARE SPLIT BY SHARED-INFORMATION SUFFICIENCY

  • 723 HEALTHCARE-S and 587 HEALTHCARE-I instances were assigned using GPT-4o at temperature 0, with correctness on shared information serving as an operational proxy for its sufficiency.The split distinguishes cases where shared information supports the correct decision from cases where additional private information is needed.Source: §D.1 · p. 29

D.2 MODEL CONFIGURATIONS

  • Evaluation covers GPT, Gemini, and Qwen model families; Claude was excluded because of API limitations.Model configurations summarize provider, version, scale, access method, and experimental usage.Source: §D.2 · p. 29

D.3 ORACLE AND STANDARD COLLABORATION

  • The oracle supplies an agent complete information as an empirical upper bound, while standard collaboration has each agent first share its own assessment.For oracle T2, the setup assumes the misleading agent is identified and uses the ground-truth context as the information state.Source: §D.3 · pp. 29–30
  • Standard collaboration shares all agents’ opinions in each round, warns that information may be misleading, and can be compared with chain or binary-tree communication; separate stages perform T1 verification, T2 state reconstruction, and T3 decision-making.A summarizer identifies the misleading agent, another agent updates the shared context from opinions and initial information, and the downstream model answers using the constructed state.Source: §D.3 · pp. 31–32

D.4 BASELINE IMPLEMENTATIONS

  • Baselines use scenario-specific prompt adaptations while keeping objectives and procedures consistent, and generally reuse the Oracle and standard-collaboration final-decision prompt.Voting independently constructs information states and decisions before majority aggregation; Robin uses a random sequential speaking order that progressively incorporates responses.Source: §D.4 · pp. 32–33
  • Manager iteratively updates and carries forward a compact information state, whereas Instructor centrally elicits targeted missing information before deciding.The Instructor prompt directs it not to invent private information or solve during discussion, and to select an agent and ask a focused question.Source: §D.4 · pp. 33–35
  • Self-Reflect lets agents choose reflection or debate, while Confidence adds peer confidence as an input signal and constrains scores to 1–100.Reflection revisits reasoning and private information; debate exchanges evidence, disagreements, and missing information.Source: §D.4 · pp. 36, 38

D.5 CONTROLLED INFORMATION-RELIABILITY SETTINGS

  • The controlled settings vary whether private or shared information is unreliable while keeping task-relevant information distributed across agents.They include reliable private and shared information, reliable private with misleading shared information, and misleading private information with reliable shared information.Source: §D.5 · pp. 38–39
  • REGROUND consistently outperforms other strategies, and performance is better with reliable than unreliable initial shared information.The evaluation focuses on unreliable private information with reliable shared information, where REGROUND is designed to identify the misleading agent and repair the collective state.Source: §D.5 · p. 39

D.6 CANDIDATE POOL SIZE ABLATION

Candidate-pool size affects reliability attribution and the downstream collaboration pipeline. Directly identifying the misleading agent improves reconstruction and task resolution, showing that attribution errors remain consequential.

D.6 CANDIDATE POOL SIZE ABLATION

  • The ablation varies M over {1, 2, 4, 5} and includes an Oracle condition that bypasses candidate selection and source attribution.Other experimental configurations remain unchanged.Source: §D.6 · p. 39
  • Smaller pools (M = 1, 2) reduce T1 and degrade T2 and T3; M = 4 is comparable to M = 3, while M = 5 degrades all three metrics.The authors suggest insufficient candidate exploration can cause the misleading source to be missed.Source: §D.6 · p. 39, Table 10 · p. 39
  • With the misleading agent identified directly, T2 rises from 0.728 to 0.813 and T3 from 0.661 to 0.837.The remaining gap indicates source-identification errors propagate to state reconstruction and downstream decisions.Source: §D.6 · p. 39

D.7 TOKEN COST ANALYSIS

ReGround uses a moderate token budget compared with discussion-intensive methods. Separately, explanation analyses link unreliable information to incorrect decisions and show how correct decisions can coexist with latent state corruption.

D.7 TOKEN COST ANALYSIS

  • 19.4K, 20.2K, and 17.6K total tokens for ReGround versus Standard’s 52.1K, 71.5K, and 31.2K across three settings with GPT-4o.The paper reports input, output, and total token use for GPT-4o and Gemini-3.5-Flash across three datasets; discussion-intensive methods use considerably more tokens.Source: Table 11 · p. 40, §D.7 · p. 40

E.1 SELECTIVE INFORMATION USE

  • 70.3% of GPT-4o explanations on HEALTHCARE-S pair unreliable information with incorrect decisions, compared with 17.4% alongside correct decisions.Across both models and datasets, unreliable information appears more often with incorrect decisions, while reliable information shows the opposite pattern.Source: §E.1 · p. 41, Figure 6 · p. 41
  • Correct task resolution can coexist with latent corruption when reasoning uses an unaffected subset of the shared information.The explanation categories are not mutually exclusive: an explanation can contain both reliable and unreliable information.Source: §E.1 · pp. 40–41

E.2 BROADER-TASK EVALUATION

The broader-task evaluation finds that misleading information remaining after a correct answer can affect subsequent clinical plans. Shared-state repair improves downstream accuracy, and additional tests examine more complex corruption patterns.

E.2 BROADER-TASK EVALUATION

  • 43% of clinical plans were judged inconsistent with ground truth despite every sampled instance being correct on its original T3 question.The plans used a misleading reconstructed state and covered diagnosis, risk, management, and next steps.Source: §E.2 · p. 42
  • Both information states answer the original query correctly, but the misleading state shifts the broader diagnosis from Prader–Willi syndrome toward type 1 diabetes and changes management.The evaluation sampled 100 healthcare instances with a correct original T3 answer and residual misleading facts, then assessed broader actionable clinical plans.Source: §E.2 · pp. 41–43, Figure 7 · p. 43

E.3 SHARED-STATE REPAIR

  • The repair test progressively replaces corrupted information with its reliable counterpart, then measures downstream task accuracy using the resulting state.Source: §E.3 · p. 43
  • Downstream accuracy generally rises as reliable information replaces corrupted information, reaching its highest or near-highest level when the state is fully reliable.The trend holds across GPT-4o, Gemini-3.5-Flash, Qwen3-32b, and three datasets.Source: Figure 8 · p. 43, §E.3 · p. 43

E.4 COMPLEX CORRUPTION STRESS TESTS

  • Stress tests use 200 HEALTHCARE instances to assess downstream decisions with three unreliable agents or misleading information distributed across agents, using GPT-4o.For the distributed-information setting, ReGround identifies misleading facts at the statement level.Source: §E.4 · p. 44
Loading 2610.01244v1…