acceptodds
Under review as a conference paper at ICLR 2027

Diagnosing and Improving Probabilistic Reasoning in Large Language Models

Abstract

Large language models (LLMs) are increasingly proposed as decision assistants who must reason probabilistically from available evidence under explicit decision costs. We propose a decision-theoretic framework that decomposes LLMs' decision loss into two components: forming accurate beliefs from provided evidence and translating those beliefs into actions that optimize a provided utility function. Using a synthetic benchmark with known ground truth, we apply the decomposition to characterize probabilistic reasoning in frontier and open-sourced models. We further evaluate whether RL interventions targeting beliefs, decisions, or both improve these components across three domains, whether improvements transfer across components and elicitation formats, and whether decision performance can improve without improvement in belief formation. We find that targeting one component of probabilistic reasoning redistributes decision loss, improving the target without necessarily transferring to others, and that jointly targeting belief formation and decision-making improves both but hinges on matched formats between training and evaluation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, probabilistic reasoning, decision-making under uncertainty, belief-action alignment

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6uXCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for diagnosing probabilistic reasoning in LLMs by decomposing total decision regret into belief loss and a residual. The framework distinguishes errors in estimating probabilities from errors in translating reported beliefs into actions under specified loss functions. The authors evaluate this decomposition across several LLM families and reasoning settings using synthetic tasks. They further investigate four RL interventions targeting belief formation, decision-making, or both. Experiments across synthetic inference, weather forecasting, and clinical reasoning show that improving beliefs does not necessarily improve decisions, while decision-only training can improve actions without substantially improving reported beliefs. Joint training improves both components, although its effectiveness depends on the training and evaluation formats.

Critical strengths
  1. The paper provides a clear decision-theoretic framework for distinguishing errors in belief formation from errors in translating beliefs into actions. The regret decomposition is simple and interpretable, and the use of controlled decision problems with known posteriors allows the authors to systematically investigate different sources of decision failures. This provides a useful perspective beyond evaluating confidence calibration or final decision accuracy alone.

  2. The paper systematically evaluates four RL interventions across three domains and two elicitation formats. An interesting finding is that improving belief accuracy does not necessarily lead to better decisions, while decision-only training can improve decision performance without comparable improvements in reported beliefs. The comparison between independent and two-turn evaluation further highlights the importance of belief-action alignment and training-evaluation format consistency. These findings offer useful insights into designing post-training objectives for probabilistic reasoning.

Critical weaknesses
  1. The decomposition relies on the model's verbalized belief to distinguish belief formation errors from decision optimization errors. However, reported beliefs may not faithfully represent the model's internal beliefs. Consequently, the optimization residual may reflect inconsistencies between elicited and internally used beliefs rather than an actual failure to optimize decisions under a given utility function. Moreover, the optimization residual is defined as a signed quantity and can be negative when the reported action performs better than the action implied by the reported belief. Therefore, improvements in the two components do not necessarily correspond to independent improvements in belief formation and decision optimization. This complicates the interpretation of the proposed diagnostic framework, particularly when attributing performance changes to specific reasoning capabilities.

  2. The post-training experiments are conducted on a single base model (Qwen3-8B), and the decision problems are restricted to binary outcomes with relatively simple threshold-based loss functions. Although the paper evaluates three domains and includes generalization to unseen thresholds, the training and evaluation tasks largely share the same underlying structures. In particular, decision-only training may learn task-specific decision rules rather than generalizable probabilistic reasoning abilities. Similarly, the strong dependence of sequential alignment training on matching elicitation formats raises concerns about whether the improvements reflect robust belief-action alignment or adaptation to a specific prompting procedure.

Questions for authors
  1. How can the authors distinguish genuine decision optimization errors from discrepancies between verbalized beliefs and the beliefs internally used by the model? Since the optimization residual can be negative, to what extent can it be interpreted as a separate source of decision-making error rather than a measure of belief-action inconsistency?

  2. The sequential alignment intervention improves performance primarily when the evaluation format matches the training procedure. Can the authors clarify whether this improvement reflects a generalizable belief-action alignment capability or simply the model learning to condition its decisions on explicitly reported probabilities?

  3. How well do the RL interventions generalize to decision problems with different structures beyond the threshold-based binary classification setting? While the experiments consider unseen thresholds, these share the same decision structure as the training tasks.

  4. Are the observed effects consistent across different base models? The diagnostic experiments include multiple model families, but all post-training experiments use Qwen3-8B. Evaluating at least one additional model would help establish whether the reported patterns are specific to the chosen architecture or initialization.

7 comments
AuthorsOpenReview ↗

W1-1 Discrepancy between verbalized and internal beliefs and how to interpret the framework

Thank you for raising the important point that discrepancies between verbalized and internal beliefs can complicate distinguishing genuine decision optimization errors from belief-action inconsistency.

Discrepancy between verbalized and internal beliefs

We wish to note that a model’s true belief is latent, thus we cannot rule out measurement errors. But the value of our framework does not depend on the reported belief being the true belief: rather, a goal of our work is to show how threats to LLMs’ decision-making can be evaluated despite a lack of access to a model’s true beliefs.

Ultimately, there is no way to separate errors in beliefs from errors in optimization without some intermediary quantity that can be measured. Our framework uses reported beliefs as this intermediary quantity, because they are easily accessible in practice and apply to both open-source and proprietary models. We evaluates whether that choice is useful by showing how the quantities in the framework can be used to drive interventions or identify patterns across models.

We currently acknowledge the potential difference between reported and true beliefs in lines 116-118 of the paper. In revision, we will further clarify why our approach does not depend on accessing true beliefs for the purpose of providing meaningful measurement of failure modes and targets for post-training.

Interpretation of belief loss and optimization residuals

We would also like to use this opportunity to clarify the interpretation of belief loss and optimization residuals.

Belief loss refers to the decision loss attributable to the reported beliefs differing from the true posterior beliefs. It summarizes several potential sources of loss before the model reports its belief, including errors in extracting information from evidence, unexpected ways of using its prior knowledge and elicitation distortion.

Similarly, optimization residual captures the decision consequences of actions being inconsistent with the optimal implied by reported beliefs. It indicates all possible changes from belief-implied to reported actions, including failures to optimize the given utility function and it using beliefs different from those they report. That is why we decided to name the term as optimization residual rather than loss.

In revising, we will add additional emphasis on the fact that these terms represent broad distinctions between error sources, which additional comparisons (as we discuss below) can extend to further disambiguate sources of loss.

AuthorsOpenReview ↗

Q1 - How the framework reflects genuine optimization errors

To directly answer your question on whether measured optimization residuals mainly reflect genuine optimization errors or belief-action misalignment, we have added a supplementary experiment.

We evaluate the decision consequences when a model is directly provided with the beliefs, and must only choose the best action. We compare three conditions: 1) independent evaluation, which elicits beliefs and actions in separate sessions; 2) two-turn evaluation, which gives the model access to evidence and its reported beliefs; and 3) belief-only evaluation, which directly supplies the reported beliefs to the model and asks for actions without the original evidence. We evaluate mean optimization residuals and agreement between belief-implied and reported actions. We test GPT-5.5 in both thinking and non-thinking modes and provide the results in the following table.

We find that with medium thinking effort, the model reliably chooses belief-implied optimal actions when probabilities are explicitly available. This suggests that in this setting, optimization residuals in independent evaluation may mainly reflect belief-action misalignment rather than a general inability to optimize under supplied probabilities. On the other hand, we find that without thinking, belief–action agreement remains between 87.0% and 89.9% across the three conditions. This may reflect genuine difficulty optimizing the given utility functions, although we cannot rule out the possibility that the model does not fully adopt the supplied beliefs.

In revision, we will explicitly clarify that our framework can be supplemented with further comparisons to tease apart more specific contributors to belief loss or optimization residuals. We will also add these results to Section 4, along with evaluations on more models.

Table 3. Changes in decision consequences when explicitly providing beliefs

Thinking modeEvaluation formatMean RoptR_\mathrm{opt}Action agreement
NoneIndependent0.01887.0%
NoneTwo-turn0.02389.9%
NoneBelief-only0.02287.7%
MediumIndependent0.004398.4%
MediumTwo-turn0.000399.8%
MediumBelief-only0.0003100%
AuthorsOpenReview ↗

W1-2 - Distinguishing independent improvements

We appreciate the reviewer raising the question on how much of the improvement in regret-based terms can be interpreted as independent improvement in separate capabilities.

It is correct that the two errors will trade off as a consequence of using the reported beliefs as the reference. We use these regret terms as our primary evaluation metrics because they quantify how interventions affect different sources of decision loss against ground truth optimal actions, which is arguably the biggest concern when relying on LLMs for decision-making.

Assessing improvements to other metrics—such as MSE for belief accuracy and agreement between belief-implied and reported actions for belief–action alignment —can certainly help corroborate our claims about the effectiveness of post-training. We have added both analyses. Overall, these secondary metrics show changes consistent with the regret-based findings, providing additional support for the observed intervention effects.

It should be noted, however, that improvements in these direct metrics will not translate into reductions in decision loss. For example, improving a belief estimate from 0.4 to 5, given the true posterior is 0.7, would improve belief accuracy but would not change the belief-implied action at a decision threshold of 0.6. The decision loss associated with that action would therefore remain unchanged, which is why we center our framework on decision consequences.

We will update the results with supplementary metrics in the revision. And we will also discuss the trade-offs between different metrics further with the added space in the revision, showing how additional metrics can be brought in to complement the information provided by our framework.

Reviewer_2ERJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a decision-theoretic framework that decomposes an LLM’s decision regret into belief loss and optimization residual, using tasks with known posteriors and explicit costs. It applies the framework to diagnose models and compare four RL post-training interventions across synthetic inference, weather forecasting, and clinical-note tasks. The interventions improve different parts of the decomposition to varying degrees, and gains depend in part on whether training and evaluation use the same elicitation format.

Critical strengths
  • The framework separates two sources of decision error. By comparing the model’s reported belief, its reported action, and the action implied by that belief under the specified costs, the paper gives a concrete way to analyze where decision regret arises.
  • The experiments test distinct interventions and evaluation formats. The paper compares belief-only, decision-only, separate multitask, and sequential alignment training, then evaluates both independent and two-turn elicitation. This supports a useful finding: improved decisions do not necessarily imply better elicited beliefs.
Critical weaknesses
  • The domain results rely on synthetic or model-generated evidence. The weather task derives labels from the HailFinder Bayesian network, and the clinical task uses SimSUM records generated from a known causal model, including AI-generated notes. These experiments support controlled comparisons, but do not establish that the same intervention effects hold for naturally collected weather or clinical data.
  • Post-training is evaluated on one base model. All intervention experiments use Qwen3-8B, so the paper does not show whether the observed training effects generalize across model families or scales.
Questions for authors

Please see the above comments.

3 comments
AuthorsOpenReview ↗

# W1- Generalization to real data settings

Thank you for the valuable comment. We wish to note that the purpose of our post-training experiments is to demonstrate how the framework can drive improvements via interventions. Without known ground truth, we cannot precisely evaluate the effectiveness of post-training interventions. We therefore focused on controlled tasks, as testing on naturally collected data introduces additional ambiguity about what should be considered ground truth. To best approximate real data, some of our tasks used more flexible inputs similar to what might be seen in LLM use in practice. This allows us to study a spectrum of tasks without losing the control that is needed to estimate sources of loss.

Across all three tasks we studied, we consistently find that belief-only and decision-only training improve their respective targets without comparable transfer to other components. Joint interventions improve both belief formation and decision-making, while these effects are more robust when the training and evaluation formats match. These findings suggest a common pattern across the evaluated tasks. To answer the question of whether similar qualitative patterns hold with naturally collected data will require validating ground truth posteriors for those settings. We will acknowledge this as important future work.

AuthorsOpenReview ↗

# W2 - Intervention effects across models

To address the concerns about replicating the intervention effects, we have added post-training experiments using Llama3.1-8B-Instruct as the base model. Overall, the intervention effects are consistent with those observed from Qwen3-8B. Belief-only and decision-only training improve their respective targets, without corresponding improvements in the other components. Separate multitask and sequential alignment training improve both components, and separate multitask training shows the most robust improvements across evaluation formats.

One difference is that, for Llama3.1-8B-Instruct, adding belief–action alignment supervision in sequential alignment training largely weakens the improvements in belief accuracy than belief-only training. For Qwen3-8B, sequential alignment training maintains a similar level of belief improvement to belief-only training. We will provide detailed results and plots in Appendix J of the revised manuscript.

Table 2-1. Intervention effects on the synthetic task, Llama3.1-8B-Instruct.

StrategyFormatΔRtotal\Delta R_{\mathrm{total}}ΔRbelief\Delta R_{\mathrm{belief}}ΔRopt\Delta R_{\mathrm{opt}}
Bindependent−0.007±0.018-0.007 \pm 0.018−0.051±0.011-0.051 \pm 0.0110.044±0.0210.044 \pm 0.021
Btwo-turn−0.064±0.013-0.064 \pm 0.013−0.050±0.010-0.050 \pm 0.010−0.014±0.011-0.014 \pm 0.011
Dindependent−0.075±0.018-0.075 \pm 0.0180.005±0.0160.005 \pm 0.016−0.080±0.026-0.080 \pm 0.026
Dtwo-turn−0.022±0.018-0.022 \pm 0.0180.006±0.0160.006 \pm 0.016−0.028±0.013-0.028 \pm 0.013
B/Dindependent−0.104±0.017-0.104 \pm 0.017−0.050±0.011-0.050 \pm 0.011−0.054±0.019-0.054 \pm 0.019
B/Dtwo-turn−0.072±0.013-0.072 \pm 0.013−0.049±0.010-0.049 \pm 0.010−0.023±0.011-0.023 \pm 0.011
B+Aindependent0.004±0.0130.004 \pm 0.013−0.012±0.011-0.012 \pm 0.0110.016±0.0170.016 \pm 0.017
B+Atwo-turn−0.021±0.014-0.021 \pm 0.014−0.010±0.011-0.010 \pm 0.011−0.010±0.012-0.010 \pm 0.012

Table 2-2. Intervention effects on the weather task, Llama3.1-8B-Instruct.

StrategyFormatΔRtotal\Delta R_{\mathrm{total}}ΔRbelief\Delta R_{\mathrm{belief}}ΔRopt\Delta R_{\mathrm{opt}}
Bindependent0.228±0.0510.228 \pm 0.051−0.466±0.051-0.466 \pm 0.0510.694±0.0810.694 \pm 0.081
Btwo-turn0.042±0.0140.042 \pm 0.014−0.053±0.014-0.053 \pm 0.0140.095±0.0240.095 \pm 0.024
Dindependent−0.126±0.013-0.126 \pm 0.013−0.002±0.005-0.002 \pm 0.005−0.124±0.014-0.124 \pm 0.014
Dtwo-turn−0.010±0.009-0.010 \pm 0.009−0.002±0.005-0.002 \pm 0.005−0.008±0.009-0.008 \pm 0.009
B/Dindependent−0.114±0.012-0.114 \pm 0.012−0.183±0.012-0.183 \pm 0.0120.069±0.0120.069 \pm 0.012
B/Dtwo-turn−0.062±0.010-0.062 \pm 0.010−0.183±0.012-0.183 \pm 0.0120.122±0.0100.122 \pm 0.010
B+Aindependent0.011±0.0080.011 \pm 0.008−0.045±0.006-0.045 \pm 0.0060.057±0.0100.057 \pm 0.010
B+Atwo-turn0.034±0.0100.034 \pm 0.010−0.045±0.006-0.045 \pm 0.0060.079±0.0100.079 \pm 0.010

Table 2-3. Intervention effects on the clinical task, Llama3.1-8B-Instruct.

StrategyFormatΔRtotal\Delta R_{\mathrm{total}}ΔRbelief\Delta R_{\mathrm{belief}}ΔRopt\Delta R_{\mathrm{opt}}
Bindependent−0.008±0.011-0.008 \pm 0.011−0.015±0.010-0.015 \pm 0.0100.007±0.0140.007 \pm 0.014
Btwo-turn−0.021±0.011-0.021 \pm 0.011−0.015±0.010-0.015 \pm 0.010−0.006±0.008-0.006 \pm 0.008
Dindependent−0.137±0.018-0.137 \pm 0.018−0.022±0.011-0.022 \pm 0.011−0.115±0.020-0.115 \pm 0.020
Dtwo-turn−0.030±0.012-0.030 \pm 0.012−0.022±0.011-0.022 \pm 0.011−0.008±0.009-0.008 \pm 0.009
B/Dindependent−0.094±0.018-0.094 \pm 0.018−0.010±0.010-0.010 \pm 0.010−0.084±0.020-0.084 \pm 0.020
B/Dtwo-turn−0.018±0.012-0.018 \pm 0.012−0.011±0.010-0.011 \pm 0.010−0.007±0.009-0.007 \pm 0.009
B+Aindependent0.002±0.0080.002 \pm 0.008−0.001±0.010-0.001 \pm 0.0100.003±0.0140.003 \pm 0.014
B+Atwo-turn−0.004±0.011-0.004 \pm 0.011−0.001±0.011-0.001 \pm 0.011−0.003±0.009-0.003 \pm 0.009

Dear Reviewer 2ERJ,

As the discussion phase is closing, we would greatly appreciate it if you could take a moment to read our responses and the revised manuscript, and please let us know if our clarification and additional experiments have addressed your concerns. We would be more than happy to provide further clarification or address any remaining questions.

Thank you!

Reviewer_pj2DOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a framework for understanding why LLMs make suboptimal decisions in settings with explicit utility functions. Specifically, the framework breaks down decision loss into two subparts: (1) belief loss: the loss incurred by holding an incorrect belief and (2) the optimization residual: the signed difference in loss incurred when taking the model’s belief-implied action vs taking the optimal action. They evaluate models using this framework on a synthetic task and then experiment with optimizing the models to minimize one or both of these losses using RL with mixed results.

Critical strengths

The paper is clearly written, and covers a consequential topic. Model performance on decision making tasks given probabilistic beliefs is not well-characterized in the existing literature to the best of my knowledge.

Critical weaknesses

The evaluation in section 4 seems fairly limited. It's not obvious to me the extent to which performance on this synthetic task maps onto performance on more realistic open domain problems.

Questions for authors

N/A

2 comments

Thank you for your constructive comments. As you noted, models’ capabilities at forming probabilistic beliefs and making decisions is an emerging but important area. In this paper, our main aim is to contribute an evaluation framework, grounded in statistical decision theory, that helps measure and improve suboptimalities in LLMs’ probabilistic beliefs and decisions.

We would like to clarify that our goal in Section 4 is to demonstrate how the framework can effectively identify different failure modes in LLM decision making, rather than to represent LLMs’ general probabilistic reasoning capabilities. We chose the controlled benchmark because it allows us to endow well-defined decision problems. Without giving the model well-defined problems with knowable ground truth, it remains ambiguous whether the LLM had a chance to arrive at the beliefs and decisions we expect it to, such that it is difficult to conclude there are losses at all.

In addition, the synthetic benchmark in Section 4 provides valuable information about what we can expert in more complex scenarios: if a model cannot optimize its behavior when provided with a well-defined decision problem, it seems unlikely to be able to form accurate beliefs and make utility-optimal decisions in more complex and uncontrolled settings. We will make this motivation clearer in the revision.

Based on your comment, we can add to the paper additional results that compare across several more realistic benchmarks where ground truth probabilities can still be established, to give a better sense of where current models struggle. The following table reports aggregate performance on three different tasks: synthetic data inference, weather forecasting, and clinical reasoning; and across two open-source models: Qwen3-8B and Llama3.1-8B. Across all three tasks, the base models exhibit nontrivial belief loss and optimization residuals, although the breakdown varies by task. We will add these additional results to the appendix to corroborate the synthetic benchmark results.

Table 1. Aggregate performance and decomposition on three different tasks

ModelTaskTotal RegretBelief LossOptimization Residual
Qwen3-8BSynthetic0.1050.0600.045
Qwen3-8BWeather0.2520.1750.077
Qwen3-8BClinical0.2890.1140.175
Llama3.1-8BSynthetic0.1370.0910.046
Llama3.1-8BWeather0.1540.224-0.071
Llama3.1-8BClinical0.2200.1110.110

Dear Reviewer pj2D,

As the discussion phase is closing, we would greatly appreciate it if you could take a moment to read our responses and the revised manuscript, and please let us know if our clarification and additional experiments have addressed your concerns. We would be more than happy to provide further clarification or address any remaining questions.

Thank you!

Comments on the paper

3 comments

We appreciate all the valuable feedback from reviewers. We are glad to see that reviewers recognized the importance of a measurement framework for LLMs’ probabilistic decision-making (reviewer pj2D); raised insightful questions on the interpretation of our framework (reviewer 6uXC); and showed great interest in the generalization of our work (reviewer pj2D, 2ERJ and 6uXC).

In response to reviewers’ comments, we have revised to make further clarification about the motivation of focusing on controlled tasks and the interpretation of the framework. We also expand the experiments to address reviewer concerns about broader performance beyond the synthetic benchmark, replicable intervention effects, and further explanations of the measured optimization residuals. We summarize the additions and clarifications below:

x\phantom{x}

Experiment choice on controlled tasks and its generalization to real data

Our main contribution is a decision-theoretic framework that decomposes sources of LLMs' decision loss. We demonstrate how to use it to diagnose LLM decision-making and how the framework can further inform and evaluate targeted post-training interventions. We choose to focus on controlled tasks because endowing decision problems with known posteriors and explicit utilities allows us to quantify the decision consequences without the added confound of not knowing what problem the model is solving.

  • To answer the concern from reviewer pj2D that our diagnosis results on the synthetic benchmark is limited to reflect LLMs’ broader performance, we would like to first clarify that the main purpose of Section 4 is to validate the measurement framework, rather than providing general claims about LLMs’ probabilistic reasoning. In addition, our synthetic diagnosis results provide valuable information about what we can expert in more complex scenarios: if a model cannot optimize its behavior when provided with a well-defined decision problem, it seems unlikely to be able to form accurate beliefs and make utility-optimal decisions in more complex and uncontrolled settings. We have updated Section 4.1 to explicitly clarify why we construct a synthetic benchmark with well-defined decision problems and further discussed its impacts in Section 6 of the revision.

    At the same time, we provide additional diagnosis results across three domains and two open-source models to broaden our empirical findings. We find that the base models exhibit nontrivial belief loss and optimization residuals across tasks, although the breakdown varies by task. We have added paragraphs in Section 5.2 to describe the findings and put the numerical results to the appendix of the revised manuscript.

x\phantom{x}

  • For reviewer 2ERJ's concern whether intervention effects hold on naturally collected data, we would like to emphasize the importance of using controlled tasks with instance-level known posteriors to evaluate the effectiveness of interventions. Using naturally collected data will introduce additional uncertainty in estimating ground truth. We agree that evaluating across a spectrum of tasks is valuable. That is why we included tasks that used more flexible inputs similar to what might be seen in LLM use in practice, including weather summary and clinical notes. We find common patterns across evaluated tasks. We have updated Section 6 to further discuss the application to naturally collected data and noted that it requires validation on estimated ground truth.

Replicating post-training interventions on Llama3.1-8B-Instruct

To address both reviewer 2ERJ and reviewer 6uXC's concerns about replicating the intervention effects, we have added post-training experiments using Llama3.1-8B-Instruct as the base model. Overall, the intervention effects are consistent with those observed from Qwen3-8B. One model-specific difference is the effectiveness of belief supervision in sequential alignment training. For Llama3.1-8B-Instruct, adding belief–action alignment supervision yields smaller improvements in belief accuracy than belief-only training. For Qwen3-8B, sequential alignment training maintains a similar level of belief improvement to belief-only training. We have added that experiment to Section 5.2 of the revision and provided detailed results in Appendix J.

x\phantom{x}

Generalization to three-action space

To answer reviewer 6uXC’s question on whether decision-only training on binary threshold decisions learns specific optimization rules or generalizable decision optimization capabilities. We test the checkpoints trained with binary decisions on a three-action space. We find that decision training can be effectively transferred to the unseen decision structures. We have added the experiment to Section 5.2 and provided full results in Appendix I of the revision. We have also acknowledged in Section 6 that this added experiment only provides successful examples of transfer, more systematic evaluation on generalizability is an important future work.

Clarifying the interpretation of the diagnosis framework

Reviewer 6uXC gives thoughtful questions that discrepancies between verbalized and internal beliefs can complicate distinguishing genuine decision optimization errors from belief-action inconsistency. We wish to first clarify that to separate errors in beliefs from errors in optimization, we have to obtain some intermediary quantity that can be measured. The observable intermediary quantity is ultimately subject to measurement errors as true belief is latent. We choose reported beliefs because of its easy access in practice and wide applicability across both open source and proprietary models. Consequently, optimization residuals defined in our framework include both optimization errors and belief-action misalignment. We have updated Section 3 to make the justification more explicit.

We acknowledge that our framework only separates broad components of decision loss from the belief or optimization side. Therefore, we have also enhanced Section 4.2 to demonstrate how to tease apart specific contributors with further comparisons. To directly answer Reviewer 6uXC’s question on whether measured optimization residuals mainly reflect genuine optimization errors or belief-action misalignment, we have added evaluation when the model is provided with explicit beliefs, and must choose the optimal action. We find that in GPT5.5, belief-action misalignment is the major source of optimization residuals in the thinking setting, while non-thinking mode exhibits a genuine difficulty in optimizing given utilities. We have provided the details and full results in Appendix K.

x\phantom{x}

Supplement metrics for testing independent improvements

To address reviewer 6uXC’s concern that regret-based terms alone cannot identify independent improvements, we have added two complementary metrics to evaluate the effectiveness of our interventions - MSE between reported and true beliefs for belief accuracy; and agreement between reported and belief-implied actions for belief-action alignment. Overall, we find that these secondary metrics show the consistent directions of improvement as the regret-based terms. We have updated Section 5.1 and Appendix E and J to include these metrics in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.