Pith. sign in

REVIEW 4 major objections 4 minor

Even frontier language models that correct medical misconceptions well in a single turn degrade sharply across multi-turn patient conversations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · grok-4.5

2026-07-15 02:35 UTC pith:DE47HFZX

load-bearing objection Useful multi-turn medical misconception setup and a clear degradation claim, but the whole result hangs on an unvalidated LLM judge and we only have the abstract. the 4 major comments →

arxiv 2607.12884 v1 pith:DE47HFZX submitted 2026-07-14 cs.CL

Evaluating Large Language Models on Misconceptions in Multi-Turn Medical Conversations

classification cs.CL
keywords multi-turn dialoguemedical misconceptionslarge language modelspatient-physician conversationserror propagationLLM-as-a-Judgemedical safetyfalse presuppositions
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that large language models cannot be trusted to keep correcting medical misconceptions once a conversation continues past the first exchange. Patients frequently embed false assumptions in their questions, and safe medical communication requires the model both to answer and to identify and repair those false beliefs as dialogue unfolds. Using a new multi-turn dataset built from real patient-physician threads, the authors find that models correcting false presuppositions at roughly 85 percent on the opening turn fall to about 50 percent within two follow-ups. An oracle experiment that substitutes correct physician answers for earlier model replies shows that much of the collapse is error propagation, yet performance remains imperfect even with clean prior context. A reader who takes the result seriously would treat single-turn medical evaluations as incomplete and would expect patient-facing systems to produce inconsistent, potentially unsafe guidance as conversations lengthen.

Core claim

Frontier models that successfully correct false medical presuppositions in a single interaction degrade substantially over subsequent turns: GPT-5 and Claude-Haiku move from roughly 85 percent correction on initial questions to about 50 percent within two follow-ups, with error propagation as a major driver and residual failures even when prior context is replaced by correct physician responses.

What carries the argument

ThReadMed-QA, a multi-turn medical dialogue dataset of 2,437 real patient-physician conversation threads (8,204 question-answer pairs) derived from AskDocs, scored by a rubric-based LLM-as-a-Judge that rates whether each response identifies and corrects misconceptions.

Load-bearing premise

The claim that a model has truly identified and corrected a misconception rests on scores from another language model applying a rubric, without reported clinician agreement confirming that the judge matches expert medical judgment.

What would settle it

Have independent physicians score the same model responses on the same multi-turn threads for whether misconceptions are identified and corrected; if they find no substantial drop from first turn to later turns, or if the LLM judge systematically disagrees with those physicians, the degradation claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Single-turn medical evaluations will systematically overstate how reliably models correct misconceptions in real use.
  • Early model mistakes can lock later turns into reinforcing rather than repairing false beliefs.
  • Even perfect prior answers do not fully restore correction rates, revealing multi-turn weaknesses beyond simple error carry-over.
  • Patient-facing deployments risk inconsistent and potentially unsafe guidance as conversations continue.
  • Evaluation frameworks for medical LLMs need to measure multi-turn behavior, not just isolated answers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Training or inference methods that explicitly re-check presuppositions at every turn may be needed to slow the observed drop.
  • Similar degradation patterns could appear in other high-stakes domains where users carry false assumptions, such as legal or financial advice.
  • Measuring agreement between the LLM judge and practicing clinicians would be a direct next test of whether the reported rates are trustworthy.
  • Longer conversations with evolving misconceptions may amplify the failure mode beyond the two-follow-up horizon already measured.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The manuscript introduces ThReadMed-QA, a multi-turn medical dialogue dataset of 2,437 patient–physician threads (8,204 QA pairs) derived from AskDocs, to evaluate whether LLMs detect and correct patient misconceptions that embed false presuppositions. Using a rubric-based LLM-as-a-Judge, the authors evaluate five models and report a consistent multi-turn degradation: frontier models (GPT-5, Claude-Haiku) correct false presuppositions at ~85% on the initial turn but fall to ~50% within two follow-ups. An oracle analysis that substitutes physician responses for prior model turns attributes much of the drop to error propagation, while residual underperformance under correct context indicates incomplete multi-turn robustness. The paper argues that single-turn medical evaluation is insufficient and that multi-turn misconception handling is a safety-relevant gap.

Significance. If the reported degradation pattern is real and the metric is valid, the work is significant for medical NLP and LLM safety: it supplies a realistic multi-turn resource grounded in authentic patient threads, documents a failure mode (error propagation plus residual context-conditioned errors) that single-turn benchmarks miss, and motivates evaluation protocols that track misconception correction over dialogue. The oracle analysis is a useful design choice for separating propagation from intrinsic multi-turn weakness. These contributions would matter for deployment guidance in patient-facing systems, provided the automated judge is shown to track clinical judgment.

major comments (4)
  1. [Abstract / Evaluation (LLM-as-a-Judge)] The central quantitative claims (~85% → ~50% correction; residual gap under oracle physician context) rest entirely on a rubric-based LLM-as-a-Judge. The abstract does not report clinician agreement, inter-judge reliability, calibration against expert labels, judge model identity, or the scoring rubric. In multi-turn medical dialogue, whether a response identifies and corrects a false presupposition is a clinical judgment; without human validation, the degradation pattern and the safety implication cannot be treated as established. This is load-bearing: either add a clinician agreement study (with agreement metrics and error analysis of judge failures) or reframe results as automated-proxy scores only.
  2. [Abstract (oracle analysis)] The oracle analysis is presented as showing that 'much of the degradation is driven by error propagation' while performance 'remains imperfect even under correct context.' Without full-text numbers (absolute rates under oracle vs. model history, confidence intervals, and how 'physician responses' are aligned to the same turns), it is unclear how much of the drop is propagation versus protocol or judge artifacts. The manuscript needs a clear decomposition table and a statement of what residual error under oracle implies for the claim of multi-turn weakness independent of self-generated history.
  3. [Abstract (ThReadMed-QA construction)] ThReadMed-QA is constructed from AskDocs threads. The abstract does not specify how misconceptions/false presuppositions are identified, labeled, or verified; whether threads were filtered for clinical seriousness; or how multi-turn 'follow-ups' are defined for evaluation (fixed protocol vs. natural thread structure). If labels or turn boundaries are model-assisted or weakly defined, the measured degradation may partly reflect dataset construction. A methods section must document labeling criteria, inter-annotator agreement if human, and the exact multi-turn evaluation protocol.
  4. [Abstract (experiments)] Only aggregate point estimates (~85%, ~50%) are given for five models, with no error bars, per-model tables, statistical tests, or stratification (e.g., by specialty, misconception type, or thread length). For a comparative empirical claim of 'consistent' multi-turn degradation, the full paper must report uncertainty and show that the pattern is not driven by a subset of threads or by judge variance.
minor comments (4)
  1. [Abstract (model list)] Model names such as 'GPT-5' and 'Claude-Haiku' should be pinned to exact API/model versions and decoding settings so results are reproducible.
  2. [Abstract (rubric)] Clarify whether 'correct false presuppositions' is a binary score or a multi-criterion rubric (identification vs. correction vs. safety of the answer), and how partial credit is handled.
  3. [Abstract (dataset)] State dataset release plans (full threads, labels, evaluation scripts) if any; multi-turn medical evaluation is hard to build on without artifacts.
  4. [Abstract (terminology)] The abstract uses both 'misconceptions' and 'false presuppositions'; define the operational relationship between these terms early in the full text.

Circularity Check

0 steps flagged

Empirical multi-turn evaluation paper; no derivation chain that forces results by construction or self-citation.

full rationale

This is an abstract-only empirical evaluation paper introducing ThReadMed-QA (2,437 real AskDocs threads) and reporting measured degradation of five LLMs on multi-turn medical misconception correction via a rubric-based LLM-as-a-Judge. There is no claimed first-principles derivation, uniqueness theorem, fitted parameter renamed as prediction, or self-definitional loop. The central numbers (~85% initial correction dropping to ~50% within two follow-ups; partial recovery under physician-oracle context) are experimental outcomes, not tautologies of the inputs. Self-citation load-bearing, ansatz smuggling, and renaming of known results are absent from the available text. The reader's and skeptic's concerns about unvalidated LLM-as-a-Judge validity are correctness/reliability risks, not circularity under the stated criteria (no quote exhibits Eq. X = Eq. Y by construction or a fitted input called a prediction). Per hard rules, honest non-finding for a self-contained empirical study against external benchmarks yields score 0 with empty steps.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

Abstract-only empirical paper. No free parameters in a fitting sense are stated. Core dependencies are domain assumptions about what counts as a misconception, the validity of LLM-as-a-Judge scoring, and the representativeness of AskDocs-derived threads. No new physical entities are invented; the main constructed artifact is the dataset and evaluation protocol.

axioms (3)
  • domain assumption A rubric-based LLM-as-a-Judge can validly score whether a model response identifies and corrects a medical misconception in multi-turn dialogue.
    Central evaluation method in the abstract; no human clinician agreement or reliability numbers are provided in the available text.
  • domain assumption AskDocs patient–physician threads are a suitable source of real multi-turn medical misconceptions for evaluating patient-facing LLM safety.
    Dataset is derived from AskDocs; generalizability to clinical care settings is assumed rather than demonstrated in the abstract.
  • ad hoc to paper Degradation measured over subsequent turns under the authors’ conversation protocol reflects multi-turn behavior rather than artifactual turn construction.
    How follow-up turns are selected, how misconceptions are annotated, and how context is fed to models are not specified in the abstract.
invented entities (1)
  • ThReadMed-QA dataset no independent evidence
    purpose: Provide multi-turn patient–physician threads with misconceptions for systematic LLM evaluation.
    Constructed resource (2,437 threads, 8,204 pairs) enabling the study; independent value depends on public release and annotation quality not shown in the abstract.

pith-pipeline@v1.1.0-grok45 · 6198 in / 2635 out tokens · 22943 ms · 2026-07-15T02:35:34.082420+00:00 · methodology

0 comments
read the original abstract

Patients seeking medical information often ask questions that embed incorrect assumptions or misconceptions. In such cases, safe medical communication requires not only answering the question, but identifying and correcting the underlying false belief. These interactions naturally unfold over multiple turns, a pattern now mirrored in interactions with LLMs. Yet current evaluation frameworks do not capture model behavior in these settings, where misconceptions can emerge, persist, or evolve over the course of a conversation. Whether LLMs can reliably correct such misconceptions over time remains largely unexamined. To study this, we introduce ThReadMed-QA, a multi-turn medical dialogue dataset of 2,437 patient-physician conversation threads comprising 8,204 question-answer pairs, derived from real patient interactions on AskDocs. This dataset enables systematic evaluation of whether models can detect and correct misconceptions under a multi-turn context. We evaluate five LLMs using a rubric-based LLM-as-a-Judge framework that scores responses based on their ability to identify and correct misconceptions. Our experiments reveal a consistent pattern: even frontier models that can address misconceptions in a single interaction degrade substantially over subsequent turns. GPT-5 and Claude-Haiku correct these false presuppositions around 85% on initial questions but drop to roughly 50% within two follow-ups. An oracle analysis replacing prior model outputs with physician responses shows that much of the degradation is driven by error propagation, while performance remains imperfect even under correct context. Even when models tend to correct misconceptions initially, their performance degrades substantially over later turns, leading to inconsistent and potentially unsafe guidance in patient-facing settings and highlighting the need for evaluation frameworks that capture multi-turn behavior.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.