acceptodds
Under review as a conference paper at ICLR 2027

Pseudo-Expert Grounding and Its Carryover in LLM Advice

Abstract

Large language models (LLMs) provide advice through conversations that establish expectations, roles, and sources of authority. We study pseudo-expert context failure (PECF) as a response-level risk pattern. Directional real-world advice in this pattern materially relies on a framework without established predictive validity for the decision. Our evaluation separates Advice Strength (A) from Grounding Source (G) and uses their conjunction as the joint PECF rate. Across four commercial models, explicitly requested BaZi advice reaches a joint rate of 92.5% in 11,512 responses. A paired probe evaluates identical follow-up questions that omit framework keywords immediately after BaZi (CP-B) or general decision-making (CP-N) histories. Across 96 scenarios, four models, and two repeats, pseudo-expert grounding occurs in 67.6% of CP-B responses and 0% of CP-N responses. The corresponding joint rates are 64.8% and 0%. On 128 stratified carryover responses, GPT-5.4 agrees with two independent human annotators on 98.4% and 97.7% of binary grounding labels. Within CP-B alone, the corresponding κ values are 0.932 and 0.901. System-prompt interventions separate changes in grounding from changes in advice frequency, with strongly model-dependent effects. Two additional-model replications extend the explicit-request finding. Together, these results establish a measurable behavioral pattern and motivate evaluating the stated basis of advice throughout a conversation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RngCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work studies PECF, where LLMs give directional advice grounded in frameworks lacking established predictive validity, such as BaZi fortune-telling. The motivation is to separate Advice Strength (A) from Grounding Source (G), because the same recommendation can rest on either evidence-based reasoning or pseudo-expert premises, and users need to know the stated basis of consequential advice. The experimental method uses four commercial models across 11,512 explicit BaZi-request responses and a paired carryover probe, scoring responses on A and G axes where levels ≥2 define the joint PECF metric. Results show explicit BaZi requests yield a high PECF while grounding carries over into follow-up questions that omit framework keywords, and system-prompt interventions separate advice frequency from grounding with strongly model-dependent effects.

Critical strengths
  • [S1] The paper studies an interesting problem: how an initial BaZi or Tarot scenario shapes the model’s responses to subsequent follow-up questions even when the follow-up omits the "BaZi" or "Tarot" keywords. The experimental results show that these pseudo-expert grounding persists into the next turn dialogue, indicating that conversational history can affect practical advice beyond the initial prompt.

  • [S2] The paper develops an A/G two-dimension rubric that scores Advice Strength (A) and Grounding Source (G) independently, with levels 0 ~ 3 on each dimension and the joint PECF metric defined as A ≥ 2 and G ≥ 2. This separation enables more fine-grained assessment and reveals how system-prompt interventions can change grounding while retaining advice frequency or vice versa.

Critical weaknesses
  • [W1] The paper investigates the tendency of LLMs to provide advice based on pseudo-expert frameworks like BaZi. However, the primary finding that models follow the requested framework when explicitly prompted is not counterintuitive. If the initial prompt explicitly asks the model to act as a traditional fortune-teller and provide a BaZi analysis, it is expected that the model would maintain that role and its associated logic in subsequent turns.

  • [W2] The presentation clarity of the paper needs improvement, and the underlying design intuition and experimental motivation require refinement. For instance, the experimental naming scheme is opaque: conditions are labeled 1C, 3A, 4A–4D, 5A, and 5B, but the paper never explains what the letters and numbers signify, making it difficult to track the design logic across tables and figures. In lines 141 and 143, the definitions for the neutral decision scenario and the fifth domain of Tarot are not clearly established. Section 3.3, the design intuition of the 4 prompts can be further expanded for comprehensiveness. Section 3.4 does not fully motivate why the two-dimensional A/G decomposition is the appropriate basis for evaluating pseudo-expert grounding. The relationship between the rubric and the binary thresholds (A ≥ 2, G ≥ 2) could be explained more transparently as well.

  • [W3] Basically all of the experimental labels are annotated by LLM judges (GPT-5.4 & MiMo-V2-Pro). The human-labeled subset comprises only 128 responses, which represents approximately 1% of the total dataset (11,512). While collecting human annotations is resource-intensive and costly, this sample size is too small to robustly validate the technical soundness of the results or support the claim in Section 4.4 that automated assessment maintains high agreement with human judgment.

Questions for authors
  • [Q1] The paper investigates PECF using BaZi and Tarot frameworks as contexts. However, the connection between these findings and the broader themes discussed in the literature is not fully explored. Specifically, how do the authors clarify how the observed PECF patterns relate to or extend existing research on safety measurement and sycophancy as mentioned in the related work section? The evaluation could be strengthened if the authors integrated more diverse or relevant scenarios to better demonstrate the practical risks of pseudo-expert grounding in LLM suggestion.
Reviewer_1L4WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper defines pseudo-expert context failure (PECF) as a pattern in which an LLM’s advice materially relies on a framework without established predictive validity, and studies it using BaZi, a Chinese fortune-telling system, as a case. Each response is scored separately on Advice Strength (A) and Grounding Source (G), and responses that cross the threshold on both are counted as joint PECF. Across 11,512 responses from four commercial models to explicit BaZi-based requests, the joint PECF rate is 92.5%, compared with 0% under non-framework control conditions. In a paired probe where the same keyword-free follow-up question is asked after either a BaZi reading or a general discussion of decision-making factors, pseudo-expert grounding appears in 67.6% and 0% of responses, respectively. System-prompt interventions shift advice frequency and grounding in different ways depending on the model. GPT-5.4 serves as the judge, with agreement against two human annotators reported on 128 carryover responses (κ = 0.90–0.93 for binary G within CP-B), and the score records and analysis scripts are released.

Critical strengths

None

Critical weaknesses

The carryover result cannot be attributed to the pseudo-expert nature of the framework.

The CP experiment, which is the paper’s headline contribution, shows that an interpretive frame set up early in a conversation keeps serving as the basis for advice on the next decision question. What it does not show is whether this is specific to pseudo-expert frameworks or just ordinary context anchoring.

(a) As the authors acknowledge in Sec. 3.2, CP-N differs from CP-B not only in the validity of the framework but also in role instructions, personalization (96 chart-specific readings per model vs. 8 generic answers), and content specificity. Acknowledging the confound does not remove it. (b) More fundamentally, the G axis only counts pseudo-expert grounding as positive. Since the CP-N history contains no pseudo-expert content to carry over, a 0% G rate in CP-N is close to guaranteed by construction. The 67.6% vs. 0% contrast therefore shows that content from the prior turn gets used as a basis for the answer, but not that it gets used because it is pseudo-expert. (c) Carryover is tested only for BaZi; Tarot and MBTI appear only in the explicit-request setting. To support the paper’s claim, I would want to see a history built on a valid personalized framework (e.g., an interpretation of the user’s health checkup results or a financial risk-profile assessment), a framework-agnostic measure of how much the answer relies on the history, and a condition that separates the fortune-teller role from the reading itself.

A second issue concerns what CP adds over 1C. Because PECF is defined by the validity of the stated basis rather than by whether the user asked for it, 1C already establishes the failure. The distinctive claim of CP must then be that the grounding persists even when it is not requested. But the follow-up comes right after a BaZi reading the user explicitly asked for, with no intervening turn, and the authors themselves note that the follow-ups can function as conversational continuations (Sec. 3.2). Omitting framework keywords is not the same as signaling a topic switch. As designed, the CP results are hard to distinguish from 1C. A condition with an explicit pivot cue (e.g., “setting the reading aside, ...”) or with unrelated intervening turns would address this.

Recent work on multi-turn context switching distinguishes pivots from refinements and treats context carryover as a model-level policy choice rather than a uniform error [1, 4]. Situating this study within that line of work, for instance as a case of “epistemic basis carryover,” and adding the controls above would make the contribution much clearer. To be clear, this concern does not affect the 1C or intervention results, but it directly weakens the carryover claim in the title.

[1] Sinha et al., Beyond Continuity: Challenges of Context Switching in Multi-Turn Dialogue with LLMs, ICBINB Workshop @ ICLR 2026.

[2] Atri et al., Evaluating Temporal Consistency in Multi-Turn Language Models, ACL 2026

[3] Gekman et al., On the Robustness of Dialogue History Representation in Conversational Question Answering: A Comprehensive Study and a New Prompt-based Method, TACL 2023

[4] Kim et al., Query-Following vs Context-Anchoring: How LLMs Handle Cross-Turn Language Switching, MME Workshop @ EACL 2026.

Questions for authors

See Critical Weaknesses

Reviewer_ck31OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper measures pseudo-expert context failures (PECF), defined as directional advice based on pseudo-expert sources. The authors do so by first asking the model to generate a BaZi reading, and then asking for some advice; they additionally measure the carryover by asking the same question, which does not mention BaZi, with two different conversation histories: a BaZi reading, or a general discussion. The experiment was conducted across four models. They then measure the strength of the advice given and if the advice relies on a pseudo-expert source (Table 2), measured via GPT-5.4. The results show that PECF occurs in the majority of responses after a BaZi reading (64.8%), as opposed to none when no BaZi reading was asked. It reaches 92.5% when the user explicitly asks for BaZi-based advice.

Critical strengths

The decomposition by advice strength and grounding is interesting and highly interpretable. The manuscript shows a careful reporting (human validation, cost…) which increases transparency.

Critical weaknesses

First, answering with BaZi is counted as a failure by definition; I understand the authors’ note on practical advice which should be separated from cultural interpretation, but I’m not sure how relevant that is in this scenario: if the user explicitly asks for BaZi, why is basing the advice on BaZi then counted as a failure? There is no measure of harm; I think this would be much more relevant if the advice harm was also measured, to see whether the PECF can actually be dangerous and how frequently. For that, I would either re-score the health and finance responses, or use scenarios where using BaZi would clearly be dangerous for the user, regardless of whether the user asked for a BaZi-based response. Addressing this weakness may change my rating.

For the carryover design experiment, the neutral controls contain no BaZi information at all, so it makes sense that the grounding results reach 0%. A more informative control could have been to include BaZi in the history, and then changing topics or explicitly asking for practical advice. Also, as MBTI is not considered pure divination, a justification as to why it is in the same category as BaZi or Tarot would be useful.

Questions for authors

Why is answering with BaZi counted as a failure even when the user explicitly asks for it? Why do the Tarot and MBTI appear only in the explicit setting and not the carryover experiment?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.