acceptodds
Under review as a conference paper at ICLR 2027

Eliciting LLM Propositional Beliefs in Agentic Tasks: Consistency, Accuracy, and Response Tendencies

Abstract

Large language models (LLMs) are increasingly deployed as autonomous agents in interactive environments that require multi-step planning, reasoning, and state tracking. Evaluating an agent's beliefs about environment states and task progress can therefore help diagnose behavioral failures. However, reference belief labels are often unavailable or difficult to obtain. In this paper, we introduce a general, LLM-assisted evaluation pipeline that combines proposition generation with reference labels. The pipeline grounds LLM-authored proposition templates in agent trajectories and obtains labels from an LLM-assisted reference observer using formal logical rules and Bayesian inference. Using this pipeline, we systematically evaluate models' self-reported belief consistency and accuracy across World, Epistemic, Affordance, and Decision propositions in ALFWorld, WebShop, and Noisy Cabinets and examine their relationship with task success. T/F-only Affordance and Decision accuracy are positively associated with task success after centering within environments. Furthermore, we find that response tendencies help explain both observed consistency and accuracy variation: the tendency to answer Unknown contributes to cross-model differences in consistency, and a True/False tendency model explains degradation in accuracy and negation consistency as context grows.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Mechanistic Interpretability, Models Belief

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CMMZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a pipeline for evaluating the accuracy and consistency of the beliefs that LLM agents hold about environment states and task progress in partially observable environments. An authoring LLM generates proposition templates and evaluation rules in four categories, namely World, Epistemic, Affordance, and Decision, which are grounded turn by turn in the agents' actual trajectories. Reference labels are computed with deterministic LLM-written rules in ALFWorld and WebShop and with a Bayesian observer in the authors' Noisy Cabinets environment, and are discretized into true, false, and unknown. The model's self-reported belief is read from the next-token probabilities of the answer letters after appending a question to the same prefix. Across six model conditions, the paper measures permutation, negation, sequential, and rephrasing consistency as well as accuracy, and claims that Affordance and Decision accuracy are positively associated with task success and that a response-tendency model explains why negation consistency and accuracy decline as the context grows.

Critical strengths

S1. The paper addresses an important evaluation gap. Prior work has mostly examined whether elicited confidence agrees with behavior and has paid less attention to whether the beliefs are actually correct. This paper queries beliefs at exactly the same prefix of the agent's own trajectory, compares them with references, and measures several consistency relations at the same time. This is a useful direction for diagnosing agent failures.

S2. The measurement design handles Unknown responses carefully. By including an Unknown option and showing how consistency and accuracy change under different treatments of Unknown answers, the paper reveals that the measurements are strongly affected by the tendency to answer Unknown. This is a useful lesson for future work on belief evaluation.

S3. Noisy Cabinets separates world truth from what can be inferred from observations. Building an environment with observation noise to distinguish world truth from references inferable from the interaction history is a serious attempt to define what the correct answer should be in belief evaluation.

Critical weaknesses

W1. The reference labels are not matched to the information available to the model, which limits what accuracy means. As the paper itself states, the observer uses information that cannot be known from the interaction history alone, such as the numerical probabilities of stochastic events and the set of possible objects. When the model answers Unknown while the reference is true or false, the answer may therefore reflect legitimate uncertainty rather than an error. The current evaluation cannot separate such cases from overly cautious answering, and the results are highly sensitive to how they are treated: AF-RL's accuracy, for example, ranges from 34.82% to 83.07% depending on whether they are counted. The authors acknowledge the observer's information advantage, but an information-matched observer that uses only the information available to the model is needed to make the accuracy numbers interpretable.

W2. There is insufficient evidence that the elicited beliefs are the ones that drive the agent's decisions. As the authors' commonsense-assertion experiment shows, the measurements change substantially merely by changing the instruction, which suggests that they partly reflect how the question is interpreted. More importantly, the paper does not check whether the reported beliefs are consistent with the agent's own behavior. A natural test would be whether the agent judges the action it actually selected to be feasible, or how often it selects actions it reports as infeasible. Without such a link, it is difficult to treat the measurements as the agent's operative beliefs rather than as answers to a separate question.

W3. The main conclusions are stated more strongly than the evidence supports. The association between accuracy and task success is a comparison among 18 model-environment conditions with R² values of 0.168 and 0.075, yet the paper describes it as "strongly predictive." The robustness checks that omit individual models or environments are helpful, but each model is evaluated on its own trajectories, so context length and the visited states differ across conditions and remain potential confounders. The response-tendency analysis shows useful statistical regularities, including the transfer of the dependence structure to held-out tasks, but the decomposition itself is an identity and does not establish the claimed fixed-response mechanism or explain why the error rates grow with context. The consistency relations themselves follow prior work such as Elazar et al., BECEL, and BeliefBank, so the contribution lies in the measurement framework rather than in new findings.

Questions for authors

Q1. How does accuracy change if the reference observer is restricted to the information available to the model? Among the current mismatches between Unknown reports and definite references, what fraction would an information-matched observer also label Unknown?

Q2. Are the self-reported beliefs consistent with the agent's own actions? For example, how often does the agent judge the action it actually selected to be feasible, and how often does it select actions it reports as infeasible?

Q3. Does the association between accuracy and task success hold in an episode-level analysis that controls for model fixed effects and context length, or in a setting where all models answer at the same prefixes?

3 comments

We sincerely thank the reviewer for the thorough, constructive review and for recognizing the same-prefix design, Unknown handling, and Noisy Cabinets.

W1 & Q1: an information-matched observer

For most of our data the references exactly match the model's information. In deterministic ALFWorld and WebShop, the rules read only the model's prompt and assign U when evidence is missing (Section 3.3). Only NC has extra priors information, and there we built the restricted observer you suggest, which relabels all five posterior-based families as U (Appendix E.1, Table 15; %):

ModelOverall: paper / restrictedU on T/F refs (ALF+Web)
Qwen3-8B43.39/44.7453.6
Qwen3-8B (think)48.82/48.8342.0
Qwen3-14B67.10/64.6113.4
Llama-3.1-8B53.61/54.7327.7
WS-RL43.56/45.0751.2
AF-RL34.82/37.3170.1

On your first question, accuracy hardly changes: Overall moves by at most 2.5 points and T/F-only by at most 0.5, and the ranking holds. On your second, 85.8% of TU/FU occur in ALFWorld + WebShop, where there is no information gap, and the restricted observer would also label only 6.8% of all TU/FU Unknown. As for AF-RL, it answers Unknown on 70.1% of the T/F propositions that its own prompt settles. Dropping NC's TU/FU lifts its accuracy only from 34.82% to 38.19%, so the gap to 83.07% reflects its tendency to answer Unknown, not the observer's priors.

W2 & Q2: elicited beliefs and the agent's own actions

As proposed, we add a test to match each prefix's Affordance report against the action then taken (Section 4.3, Table 4; Appendix E.5, %):

ModelChosen action reported feasibleReported infeasible, then chosen
Qwen3-8B78.91.3
Qwen3-8B (think)98.90.0
Qwen3-14B79.51.3
Llama-3.1-8B97.80.4
WS-RL80.50.6
AF-RL70.01.3

Models judge the action they then take feasible in 70.0–98.9% of cases, and only 0.0–1.3% of the actions they report infeasible are then taken.

On the commonsense experiment, the two prompts differ in the evidence the model may use. Under the interaction-only prompt, U is the correct answer to a fact absent from the interaction, so the change shows models following the stated scope. All our evaluations use one fixed prompt, under which reports match actions.

W3 (a): claim strength

By "strongly predictive" we meant that the association is statistically robust (95% CIs > 0 in Figure 3; Table 16), not that R2R^2 is high. We modify our results and it now says "general correlates of task success."

W3 (b) & Q3: confounders, episode level, and same prefixes

  • Confounders (Table 17): Both correlations keep 95% CIs above zero with standardized turn bins, fixed retention, or Overall accuracy, so trajectory length, T/F retention, and the T/F filter do not explain them.
  • Episode level (Figs. 10, 17): Within-condition bins and task + model×environment fixed effects remove model and environment differences; both slopes stay positive (CI above zero for Affordance on new models).
  • Same prefixes (Appendix G.2, Table 25): Shared prefixes reproduce our consistency and response-tendency results; task-success regressions require own trajectories.

W3 (c): decomposition and the response-tendency model

Our claim is the response-tendency model: the errors that grow with context attach to statement polarity. Appendix F.6 (Tables 22–24) fits fact error (flips all reports on ϕ\phi), response tendency (one polarity), and random error (per reply) to accuracy changes alone, predicting first-to-last-bin agreement changes (pp, Table 23):

RelationObservedFact errorResponse tendencyRandom error
Permutation−0.2 [−0.7, +0.2]0+1.0−6.3
Rephrasing (ϕ\phi false)−3.6 [−11.5, +3.6]0+3.0−14.7
Negation (ϕ\phi true)−5.7 [−7.9, −3.5]0−4.0+2.0
Negation (ϕ\phi false)−6.7 [−9.0, −4.2]0−12.3−11.5

A fact error model predicts no change, a random error model predicts permutation and rephrasing drops that do not occur, only the response tendency model we made fits the data. And we agree that why the errors grow with context is a separate causal question that needs interventions (Appendix H).

W3 (d): contribution

The consistency relations do follow prior work (Section 4.1); we use them as instruments. However, three findings are new: Unknown tendencies confound cross-model comparisons (Section 4.2), Affordance and Decision accuracy are general correlates of task success (Section 4.3), and response tendencies link context-length accuracy drops to negation inconsistency and transfer to held-out tasks (Section 5).

We thank the reviewer again and welcome further discussion.

Thank you for the thorough response and the additional analyses. They address most of my concerns, and I have raised my score to 3. I encourage the authors to include these analyses in the revised paper.

Thank you for your follow-up and for recognizing our rebuttal efforts. We truly appreciate your time and the positive update. We have already include the new analysis, tables and figures into our revised manuscript and will further integrate all of them into the future camera-ready version.

Reviewer_fTAROpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates the accuracy, logical consistency, and behavioral relevance of Large Language Model agents’ propositional beliefs in interactive tasks. To address the difficulty of obtaining reference belief labels in observable environment, the papers propose an automated LLM-assisted belief evaluation pipeline. The framework generates proposition templates and formal evaluation rules, grounds them in agent trajectories across four categories, and derives reference labels using a rational obserer based Bayesian inference and logical rules.

The main contributions are threefold: (1) introducing a normative framework to evaluate propositional beliefs in partially observable environments; (2) examining the ralationship between belief accuracy and downstream task success; (3) To explain how preferences for “unknown” responses and asymmetric responses to affirmative and negated propositions affect measured belief accuracy and logical consistency, developing a response-tendency model.

The experimental results show that the accuracy of Affordance and Decision beliefs is positively correlated with task success, whereas World and Epistemic belief accuracy exhibits little correlation. The findings also indicate that models’ tendencies to respond “unknown” substantially influence consistency measurements. Furthermore, belief accuracy and negation consistency generally deteriorate as interaction contexts become longer.

Critical strengths
  1. A structured and extensible proposition generation pipeline: It separates proposition template generation from formal evaluation rule construction, and grounds them in environment information and agent trajectories. This modular design enables diverse, evaluable propositions across environments and reduces reliance on task-specific formulations.
  2. A rational reference observer for belief evaluation under partial observability: Built on logical rules and Bayesian inference, the observer's information set lies between the agent's available information and the complete world state, allowing evaluation to account for uncertainty rather than relying solely on physical world truth, thus providing a principled reference for belief accuracy.
  3. Comprehensive belief evaluation and analysis of response tendencies: It evaluates agents' beliefs in terms of accuracy and four consistency dimensions, examines the relationship between belief accuracy and task success, and analyzes "Unknown" response tendencies with a response-tendency model to explain variations in belief accuracy and logical consistency, enhancing the framework's diagnostic value.
Critical weaknesses

1.The paper uses self-reported beliefs as proxies for agents internal beliefs. Although the authors demonstrate correlations between belief accuracy and task success, they do not directly examine whether these self-reports align with the beliefs that agents rely on duringaction selection. This limits the evidence supporting the behavioral relevance of the proposed belief measurements.

2.The belief accuracy-task success analysis relies on T/F only subsets whose retention rates bary substantially across model. Since Unknown response tendencies are themselves model-dependent, it remains unclear how much the reported associations are affected by response-dependent sample selection rather than differences in belief accuracy alone.

Questions for authors

1.Could the authors examine whether self-reported Affordance and Decision beliefs are consistent with agent’s actual actions? In particular, how frequently do agents select actions that contradict their reported beliefs , and are such inconsistencies associated with task failures?

2.The reference observer has access to certain environmental information unavailable to theagent, as discussed in Section 3.3 and Appendix C. Could the authors quantify how this information asymmetry affects reference labels and measured belief accuracy? For example, would the results remain similar under alternative observer priors or morerestricted information assumptions?

3.Given the substantial cross-model differences in Unknown response tendencies, could theauthors examine whether the positive correlations between T/F-only Affordance/Decisionaccuracy and task success remain robust when controlling for query retention rates orusing a common evaluation subset?

3 comments

We sincerely thank the reviewer for the thoughtful review and the positive assessment of the pipeline, reference observer, and response-tendency analysis.

W1 & Q1: consistency between self-reports and the agent's own actions

We agree this is important to test behavioral relevance. As suggested, we added a direct comparison between self-reported beliefs and subsequent actions (Section 4.3, new Table 4; Appendix E.5), matching each prefix's action-feasibility Affordance propositions (e.g., "You could open the drawer 1 right now") against the action actually taken:

ModelAction taken reported possible (%)Reported not possible, then taken (%)
Qwen3-8B78.91.3
Qwen3-8B (think)98.90.0
Qwen3-14B79.51.3
Llama-3.1-8B97.80.4
WS-RL80.50.6
AF-RL70.01.3

Each model reports the action it then takes as possible in most cases. The remaining answers are largely Unknown for the models with a high Unknown rate (Section 4.2), and an Unknown report does not contradict the action. The more direct test: how often a model takes an action that it has just reported as not possible, happens for at most 1.3% of such reports.

You also asked whether such inconsistencies come with task failure. If failures arose from reports diverging from the beliefs behind the action, conflicts would be more frequent in failed episodes. However, our model doesn't show this (Appendix E.5).

These results indicate that self-reports align closely with the beliefs guiding action selection.

Q2: effect of the observer's information advantage

By carefully checking our code, the observer hold an information advantage only in NC: in deterministic ALFWorld and WebShop, reference rules use only the model's own prompt context and assign U whenever information is missing (clarified in Section 3.3). To measure the effect of NC's extra priors (noise rates and world structure), we rescored all models under a conservative restricted observer that relabels all five posterior-based NC proposition families (6.4% of queries) as U, and on ALFWorld + WebShop alone (%):

ModelOverall: paperrestricted observerALFWorld + WebShopT/F-only: paperrestricted observerALFWorld + WebShop
Qwen3-8B43.3944.7442.5076.1876.0976.72
Qwen3-8B (think)48.8248.8348.0176.4276.1477.33
Qwen3-14B67.1064.6166.1581.0680.6280.89
Llama-3.1-8B53.6154.7353.0269.5969.4269.11
WS-RL43.5645.0742.8075.4875.2076.00
AF-RL34.8237.3133.2477.1976.7076.95

The results are stable (Table 15): under the restricted observer, accuracy moves by at most 2.5 (Overall) and 0.5 (T/F-only) points and model rankings are unchanged. Accuracy–success correlations also stay positive (Table 17, rows 5–6 below), so the findings do not rely on the observer knowing more than the model.

W2 & Q3: T/F-only selection and retention rates

Figure 3 uses T/F-only accuracy on purpose, to keep belief accuracy separate from how often each model answers Unknown (Section 4.2). As suggested, we add more controled regression results by repeating the regression with one possible source of selection removed (rr [95% CI]):

Accuracy and controlAffordanceDecision
1. T/F-only (Figure 3)0.41 [0.27, 0.52]0.27 [0.14, 0.38]
2. Overall accuracy (all queries, U included)0.21 [0.08, 0.31]0.55 [0.44, 0.63]
3. T/F-only, retention rate held fixed (partial rr)0.42 [0.29, 0.53]0.16 [0.02, 0.28]
4. T/F-only, standardized over turn bins0.33 [0.17, 0.45]0.18 [0.05, 0.30]
5. T/F-only, restricted observer0.41 [0.27, 0.52]0.35 [0.20, 0.47]
6. T/F-only, without NC0.45 [0.33, 0.56]0.36 [0.24, 0.47]
  • Response-dependent selection (row 2): Both correlations keep CIs above zero when considering Unknown labels, so the T/F-only filter does not affect our results.
  • Retention rate (row 3): If models that keep more T/F answers simply scored better, fixing the retention rate would remove the association; the partial correlations stay positive. Rows 4–6 show the same for turn composition and the NC reference.

To add more strength to our results, we collect some new data on shared prefixes where every model answers the same queries. Our consistency and response-tendency findings are reproduced. However, sine task success is a trajectory-level metrics, we cannot do accuracy–success regressions on same-prefix sets.

We thank the reviewer again and welcome further discussion.

Thanks for your response.

Thank you for your time and the positive update.

Reviewer_XPzvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces a benchmark in which LLM agents perform multi-step tasks and agents’ self-reported beliefs are queried at intermediate points through the tasks. Beliefs are elicited about the correctness of four types of propositions about the task environment: World propositions, epistemic propositions, affordance propositions, and decision propositions. Correlations between the accuracy of self-reported beliefs and task success are reported, showing the belief types for which accurate beliefs appear important for success. In the tasks studied in this work, affordance and decision belief accuracy correlates with success, but world and epistemic accuracy do not. The work also studies the consistency of self-reported beliefs.

Critical strengths

The topic of LLM self-reported belief accuracy, and how self-reported belief corresponds to success, is both interesting and important for understanding agent behaviors. The evaluations are well-designed and produce meaningful and interesting results. The paper is overall well written.

Critical weaknesses
  • The authors evaluate only small LLMs (14B parameters and under). I wouldn’t be surprised if the results in the paper do not generalize to larger (and more relevant) LLMs. 5 of the 6 model variants are in the Qwen family, which makes it unclear how results would vary across model families.

  • The result that belief types that correlate with success (Decision and Affordance, but not World or Epistemic) could be largely driven by task selection. The authors should discuss this and discuss how features of the selected tasks may influence their results.

  • “Reference labels” and “proposition templates” should be defined when the terms are first used. I found the term “reference label” especially confusing, because it was unclear that this is defined in terms of the reference observer.

  • The number of tasks is not reported as far as I can see. This is a critical omission, because if the number is very low then the results may be unreliable.

Questions for authors
  • Why were World and Epistemic accuracy uncorrelated with task success? Is this simply due to the particular tasks in this evaluation not requiring World and Epistemic accuracy, or is it a more general result about agents? I could imagine tasks where World and Epistemic accuracy would be critical.

  • Why were only small LLMs (14B and under) chosen, and why were 5 of 6 LLMs Qwen models? Is there a reason why evaluating larger LLMs is impractical?

4 comments

We sincerely thank the reviewer for the careful reading and the encouraging comments on the topic, the evaluation design, and the writing.

W1 & Q2: model size and model families

We kept to small models for two practical reasons. First, every proposition is queried under all 6 option orders and in several variants at every turn, which takes millions of evaluations. Second, stronger models are close to ceiling on belief accuracy and task success in ALFWorld and NC, which leaves little variation to analyze.

To show the robustness of our paper's main results, we added four models from four other families on a stratified one-third task subset (Appendix G, Table 25):

ModelSizeTask success: ALF / Web / NC (%)T/F-only neg / acc, own (%)Same, shared prefixes (%)
DeepSeek-V4.1-Flash552B88.3 / 45.8 / 97.597.3 / 97.797.2 / 97.8
Gemma-4-31B31B64.2 / 39.2 / 95.097.5 / 97.296.7 / 97.3
gpt-oss-20b21B48.3 / 17.5 / 80.090.2 / 91.390.5 / 92.4
Phi-414B14.2 / 11.7 / 87.582.8 / 85.781.9 / 83.9

The paper's conclusions remain robust across these models (Appendix G.1–G.2):

  • Unknown tendency (Section 4.2): Almost all Unknown answers to commonsense assertions disappear once the prompt allows general knowledge (Figures 16, 19). As in Figure 2, Unknown answers follow the prompt's evidence scope, not missing knowledge.
  • Task success (Section 4.3): The Figure 3 regression again gives positive Affordance and Decision correlations (r=0.47r=0.47, 0.610.61; 95% CIs above zero; Appendix G.1), and within a condition, success rises with trajectory accuracy (Figure 17).
  • Response tendencies (Section 5): Observed contradiction rates track the rates predicted from accuracy alone and lie below y=xy=x (κ<0\kappa<0), and negation agreement and accuracy decline together with context length (Figures 18, 20), as the response-tendency model explains.

W2 & Q1: role of task selection for World and Epistemic accuracy

We agree that World and Epistemic accuracy matters for certain task. Our intended message is that Affordance and Decision accuracy are the more general correlates of task success. We add a new Table 3 to break down the Figure 3 regression (rr [95% CI]) by environment and on task-relevant propositions mentioning goal entities:

QueriesWorldEpistemicAffordanceDecision
All (Figure 3)0.01 [−0.11, 0.13]−0.02 [−0.13, 0.10]0.41 [0.27, 0.52]0.27 [0.14, 0.38]
ALFWorld0.38 [0.27, 0.48]0.33 [0.20, 0.45]0.53 [0.37, 0.66]0.53 [0.24, 0.74]
WebShop−0.20 [−0.42, 0.03]−0.10 [−0.30, 0.09]0.28 [0.10, 0.44]0.44 [0.30, 0.56]
NC−0.46 [−0.66, −0.18]−0.76 [−0.89, −0.46]0.35 [0.06, 0.59]0.09 [−0.27, 0.42]
Task-relevant0.24 [0.11, 0.35]0.22 [0.04, 0.38]0.57 [0.39, 0.68]0.29 [0.14, 0.40]
  • Affordance & Decision are positive in every environment and on task-relevant propositions, so the pooled correlation does not depend on which environments or propositions are selected: a general relationship.
  • World & Epistemic meanwhile is near zero because environments point in opposite directions (positive in ALFWorld, near zero in WebShop, negative in NC). Their relation to success depends on task relevance.

W3: definitions of "proposition template" and "reference label"

Thank you for pointing this out. We now define each where it first appears:

  • Proposition template (Section 3.2): An LLM-generated statement template for an environment (e.g., whether an object is at a location), filled with task and context entities to form grounded propositions (e.g., "The mug 1 is currently in/on the stoveburner 1").
  • Reference label (Section 3.3): The hard True/False/Unknown label from the reference observer—assigned directly by logic rules in deterministic ALFWorld and WebShop, or by thresholding posterior probabilities (0.05 / 0.95) in NC.

W4: number of tasks

We now report these numbers in Section 3.2 and Appendix A.1 (new Table 5, which also gives per-condition success rates). The sample is large and stratified: tasks are drawn across categories in ALFWorld (6 × 60) and WebShop (5 × 72), plus 120 NC worlds, and each of the 6 model conditions runs every task once.

ALFWorldWebShopNCTotal
Tasks360360120840
Trajectories2,1602,1607205,040
Turns (all queried)63,49314,92011,48089,893

We thank the reviewer again and welcome further discussion.

Thank you for your detailed response. The large number of trajectories makes it reasonable to only include smaller LLMs, and I appreciate the more varied range of LLMs you now test.

On NC, why do you think World and Epistemic accuracy is negatively correlated with task performance? Is there a reason or mechanism that one would expect to apply to real-world tasks, or is this a peculiarity of the tasks in NC?

Thank you for your follow-up question. The negative correlation observed between task success and World/Epistemic belief accuracy in NC stems from the characteristics of partially observable/non-deterministic environments, driven by two key mechanisms:

  1. Conservative / "Unknown" Bias: In stochastic or partially observable domains like NC, ground-truth reference labels for World and Epistemic states are frequently Unknown. We observed that weaker models (with lower task success) exhibit a strong conservative bias, frequently defaulting to Unknown without active reasoning. Consequently, they achieve high apparent belief accuracy simply by matching the Unknown reference distribution.
  2. Selection Bias under TF-filtering & Variance in Inference Complexity: When filtering out Unknown to evaluate TF-only accuracy, the resulting query distribution shifts across models. Unlike Affordance and Decision queries, where subcategory difficulties are relatively uniform, World and Epistemic subcategories vary drastically in posterior reasoning complexity. Models with a high Unknown rate filter out difficult queries, leaving primarily simpler ones. Conversely, stronger models (e.g., Qwen3-14B) attempt definitive True/False commitments on challenging inference queries (such as known location in Table 7), retaining a much harder query distribution in TF-only, which penalizes their measured accuracy.

The first effect could recur in other stochastic environments: wherever weaker agents conservatively answer on hard World and Epistemic questions, accuracy will fail to forecast the task success. This supports our conclusion that World and Epistemic accuracy are not robust predictors of task success. Affordance and Decision accuracy, which bear directly on the agent's next action (Table 4), are more robust correlates

Thank you for your response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.