{"id":"29a08f3c-348e-491b-9650-416d35ac5c0d","arxiv_id":"2507.02197","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LLM role-playing agents show systematic belief-behavior inconsistencies in a Trust Game, detectable before simulation via prompt-based belief elicitation.","lead":"This paper tests whether LLM role-playing agents act consistently with the beliefs they state about the people they simulate. It finds systematic mismatches between what the models say and what they do in a trust game, and proposes a framework to check this before generating synthetic data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central 'inconsistency' metric compares different tasks—belief-ranking or estimation prompts versus role-play decisions—so the headline finding may be a prompt-format artifact rather than a genuine belief-behavior gap.","rationale":"I agree with the reader's weakest_assumption: this is the most load-bearing concern because it attacks the construct validity of every consistency metric in Section 4 and the forecast measure in Section 5. The paper does provide useful evidence for its framework—multiple models, belief-elicitation strategies, both population- and individual-level analyses, and a candid limitations section. However, the central claim is not settled without a matched-prompt control. The paper also has additional issues, notably the Section 4.2 claim about CTX+$ contradicting Table 3 for Llama 3.1 8B, and Appendix D.1 Table 9 referencing human data that is never described in the paper. I mention these in passing, but the prompt-equivalence problem is primary: if the belief and behavior prompts activate different task priors, the measured 'inconsistency' could disappear when the prompts are aligned. The verdict remains CONDITIONAL because the framework and the broad qualitative finding are plausible and addressable in revision, but the specific quantitative conclusions should not be accepted until the equivalence check is run.","tokens_in":21072,"tokens_out":6106,"duration_ms":74222,"concrete_test":"Run a matched-prompt control on the same N=50 personas and all models/strategies: take the exact role-play prompt (App. B.1, or App. B.2 for multi-round) and change only the final instruction from 'How much will you transfer?' to 'Without acting, predict the exact dollar amount you would transfer in this identical situation,' keeping all other wording, context, and output format identical. Compute the same rho and MAE between predicted and enacted amounts, with multiple seeds and bootstrap confidence intervals. If predicted and enacted amounts are near-identical under matched prompts, the original inconsistency is a prompt-format artifact; if substantial mismatches persist, the belief-behavior gap is genuine.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Population-level 'beliefs' (App. C.1.1/C.1.2) ask the model to rank trait levels by 'interpersonal trust'; CTX+$ asks for mean/SD dollar estimates; the role-play prompt (App. B.1) asks for a single transfer amount with game instructions and a BELIEF/DESIRE/INTENTION framing. Section 4.1 then scores consistency as Spearman rho and Delta-eta-squared between these outputs. But a rank over a latent construct and an action in a strategic game are not interchangeable probes. Trust Game transfers reflect risk preferences, reciprocity, and forward-looking motives; the sample explanations in App. C.4 show the model reasoning strategically (e.g., sending $0 to preserve later rounds). A model can therefore hold a stable belief about 'trust' and still choose a low transfer, making the measured divergence an artifact of construct/task mismatch. The same issue affects individual-level forecasts: the forecast prompt (App. C.3.1) asks 'How much will you send?', while the enacted run uses ReAct (App. B.2) with different instructions and history. The paper nowhere validates that these prompts measure the same latent quantity. Without such validation, the central claim of systematic belief-behavior inconsistency is not established; the result may instead be prompt sensitivity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an evaluation framework for belief-behavior consistency in LLM-based role-playing agents, using an augmented GenAgents persona bank and the Trust Game as a testbed. It elicits population-level beliefs about how persona attributes affect trust decisions under three prompting strategies (NOCTX+TR, CTX+TR, CTX+$), measures consistency against simulated transfer behavior via Spearman correlation and eta-squared discrepancy, and tests self-conditioning and imposed priors. At the individual level, it asks a model to forecast its own future send amounts across six rounds against fixed Trustee archetypes and compares forecasts with enacted ReAct-based decisions. The central claim is that LLM agents exhibit systematic belief-behavior inconsistencies, that in-context conditioning is unreliable for controllability, and that forecasting accuracy degrades over longer horizons.","tokens_in":21216,"tokens_out":8429,"duration_ms":92124,"significance":"If established, the framework would be a valuable pre-hoc diagnostic for synthetic behavioral data, extending post-hoc evaluation benchmarks with population- and individual-level consistency metrics, effect sizes, multi-turn analysis, and an explicit controllability test. The paper uses multiple open-weight LLMs, reports quantitative metrics rather than LLM-based judgment, and states that code and prompts will be released. However, the current evidence does not yet support the strength of the central claims: the quantitative support is thin (N=50 personas, no confidence intervals or significance tests), and two load-bearing design choices—construct alignment between belief prompts and behavioral outcomes, and information equivalence between forecast and enactment prompts—are not validated. These issues are potentially fixable, so the work is a promising direction rather than a definitive demonstration.","major_comments":[{"comment":"The population-level consistency metric treats divergence between a ranking of 'interpersonal trust' (NOCTX+TR, CTX+TR) and observed dollar transfers in the Trust Game as 'belief-behavior inconsistency.' These are different constructs: transfers in the Trust Game reflect risk preferences, strategic foresight, and reciprocity expectations, not only interpersonal trust. The model's own reasoning in Appendix C.4 (e.g., Table 8, round 1: sending $0 to 'err on the side of caution') shows that a stable belief about trust need not imply a high transfer. The paper does not validate that the belief-elicitation prompts and the role-play prompt probe the same latent quantity. Without such validation, the measured inconsistency may be a prompt-format or construct artifact. The CTX+$ condition, which asks for dollar estimates, yields different results (Table 3), consistent with this concern. The authors should either establish equivalence empirically (e.g., by comparing trust-rankings with dollar-based rankings from the same model) or reframe the headline as cross-prompt sensitivity rather than belief-behavior inconsistency.","section":"Section 4.1, Table 2"},{"comment":"The individual-level forecast comparison is confounded by information asymmetry. The forecast prompt in Appendix C.3.1 asks for a send amount given only the round number, the opponent archetype, and a generic reference to 'the history of the game so far,' but it does not include the actual previous rounds' sends and returns. The enacted ReAct role-play prompt in Appendix B.2 explicitly includes a 'GAME HISTORY' section with the actual round history. Thus the forecast and enactment conditions are not informationally equivalent: the forecaster is asked to predict without the accumulated history that the actor sees. The observed increase in MAE over rounds (Fig. 1) could simply reflect this missing information rather than a genuine degradation of belief-behavior consistency with horizon. The authors must either provide the same history in both prompts or explicitly frame the forecast as a from-scratch state-based prediction and interpret the results accordingly.","section":"Section 5.2, Appendices B.2 and C.3.1"},{"comment":"The quantitative support for the central 'systematic inconsistency' claim is weak. Table 3 reports median Spearman correlations and |Δη²| values computed over only 9 attributes for N=50 personas, with no confidence intervals, significance tests, or multiple-comparison correction across the 9×3×3 attribute-strategy-model cells. Table 4 reports a single Spearman value per condition with no measure of variability, so claims such as 'self-conditioning enhances alignment in Llama models' may be within sampling noise. The individual-level analysis (Section 5.3) is based on a single persona (App. C.4) and six rounds, and the R² values in Fig. 1 (e.g., 0.04 for Llama 3.1 8B under M3) show that some series are essentially flat, weakening the 'near-monotonic' wording. The authors should report bootstrap confidence intervals, permutation tests, or at least per-persona variability for the individual-level results.","section":"Tables 3 and 4, Section 4.2"},{"comment":"The perturbation procedure for imposed priors is not specified. The text says weak and strong perturbed priors are constructed to have ρ=0.80 and ρ=0.20 with the original elicited beliefs, but no algorithm is given for how a prior with a target Spearman correlation is generated, which attributes are flipped, or how ties are handled. Without this detail, the controllability results in Table 4 are not reproducible, and the interpretation (that imposed priors undermine consistency) may be an artifact of the specific perturbation scheme chosen. The authors should provide the exact procedure and the actual prior statements used for each model and attribute.","section":"Section 4.3, Appendix C.2.3"}],"minor_comments":[{"comment":"The sentence 'The Trust Game offers quantifies interpersonal trust as the amount...' contains a grammatical error ('offers quantifies'); it should be 'quantifies' or 'offers a quantification of.'","section":"Section 1"},{"comment":"The observation about Gemma 2 27B's ρ dropping under self-conditioning is stated twice in the same paragraph; one of the sentences should be removed.","section":"Section 4.3.1"},{"comment":"The claim that supplying Trust Game context 'failed to increase Spearman rank correlation... across all LLMs and attributes' is not true at the attribute level: for example, in Table 3 the Age row for Llama 3.1 70B shows ρ=-1.00 under NOCTX+TR and ρ=0.80 under CTX+TR. The claim should be qualified to the median or tested statistically.","section":"Section 4.2"},{"comment":"The caption of Table 9 says 'Mean Spearman correlations between model predictions and human trust behavior,' but the experiment appears to compare model predictions to simulated role-play behavior, not human data. The caption should be corrected to avoid implying a human-subject validation that the paper does not contain.","section":"Appendix D.1, Table 9"},{"comment":"The population-level role-playing prompt includes a placeholder '$theoretical_props' that is not defined in the main text. If this field is non-empty in the 'unconditioned' baseline, the baseline is not condition-free; if it is empty, this should be stated explicitly for replicability.","section":"Appendix B.1"},{"comment":"The paper should report the size of the test split of the GENAGENTS persona bank and the number of personas per attribute level, since N=50 is small and some attribute levels may have very few personas, making the eta-squared estimates unstable.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and important question, and the proposed framework has genuine potential as a pre-hoc diagnostic. However, the two core validity threats—construct mismatch in the population-level metric and information asymmetry in the individual-level forecast design—undermine the central interpretation as currently presented. Both are fixable with targeted experiments (adding history to forecast prompts, eliciting dollar-based beliefs in all conditions, reporting uncertainty). I would not reject: the contribution is substantial and the issues are within the scope of a major revision. I would also encourage the authors to tone down the 'systematic' claim until statistical support is provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look. The paper's core idea—elicit beliefs before running a role-play simulation and use the gap as a pre-hoc validity check—is genuinely useful, and the two-level (population and individual) Trust Game setup is a clean testbed. The conditioning experiments (self vs. imposed priors) are a nice addition; they act inside the role-play context and give the paper value even if the main comparison is challenged. Table 1 is a fair map of the evaluation space.\n\nBut the stress-test note is on target. The headline inconsistency metric compares different tasks. NOCTX+TR asks for a ranking of 'interpersonal trust'; CTX+$ asks for mean and SD dollars; the role-play prompt asks for one transfer amount with game rules and a BELIEF/DESIRE/INTENTION frame. A rank over a latent construct is not an interchangeable probe of a decision in a strategic game. The model's own explanations show strategic reasoning (sending $0 to preserve later rounds), so a stable belief about trust can coexist with a low transfer. Without validating that the prompts measure the same latent quantity, the measured 'inconsistency' is mostly prompt sensitivity. That is a load-bearing flaw, not a quibble.\n\nThere are also smaller problems. N=50 personas with nine attributes and no confidence intervals or multiple-testing correction makes the population-level numbers more illustrative than quantitative. Section 4.2's claim that CTX+$ gives better rank consistency is contradicted by Table 3 for Llama 3.1 8B (rho 0.00 vs 0.40 and 0.80 for the construct strategies). Appendix D.1, Table 9, refers to 'human trust behavior' but no human data appears anywhere; that needs to be either supplied or reworded.\n\nWhat survives? The framework, the conditioning results, and the qualitative point that self-reported beliefs often don't predict later actions in LLM agents. Those are worth publishing after the metric problem is addressed. As it stands, the paper overstates its central finding.\n\nWho is this for? People building or evaluating LLM role-playing agents for behavioral simulation. I'd send it to peer review, but with a clear request: validate or narrow the belief-behavior metric, add basic inference or drop the quantitative claims, and clean up the internal contradictions. Right now it's a conditional accept story, not a headline result.","headline":"Good framework, unvalidated yardstick: the headline belief-behavior gap is measured across different prompt formats, so the paper's main number should be read as prompt sensitivity until equivalence is shown.","tokens_in":21822,"tokens_out":3139,"would_cite":true,"duration_ms":34108,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LLM role-playing agents systematically fail to act on their stated beliefs, and a new evaluation framework can measure that failure before a simulation runs.","keywords":["belief-behavior consistency","role-playing agents","LLM simulation","Trust Game","synthetic human behavior","belief elicitation","forecast horizon","persona consistency"],"falsifier":"Reword or counterbalance the two prompt formats: for example, ask the belief-elicitation question in role-play style and the role-play question in belief style, or administer several semantically equivalent variants of the belief prompt. If belief-behavior consistency changes substantially across superficial rewrites, the paper's central gap is at least partly a wording artifact; if it survives, the gap is robust.","tokens_in":20804,"feed_emoji":"🎭","tokens_out":5578,"duration_ms":65910,"temperature":0.7,"pith_summary":"This paper tries to establish that LLM-based role-playing agents are systematically incoherent: the beliefs they state when asked directly do not reliably match the actions they take when playing a role. Using an augmented persona bank and the Trust Game, it builds a belief-behavior consistency framework that works at two levels, across whole simulated populations and for single agents forecasting their own future moves. It then asks whether giving game context during belief elicitation helps, whether asking about dollar amounts rather than abstract trust helps, and whether longer forecasting horizons hurt. The paper finds that context does not fix inconsistency, dollar-targeted beliefs improve rank ordering but inflate effect sizes, self-conditioning helps some models but not others, imposed researcher priors generally break alignment, and forecasting degrades as the horizon grows. If true, belief elicitation becomes a cheap diagnostic for synthetic-data validity rather than something to trust as a description of what the simulation will do.","feed_headline":"LLM role-players don't act on their stated beliefs","feed_subtitle":"A Trust Game study measures the gap between what agents say and what they simulate doing, before data generation.","key_machinery":"The load-bearing object is a two-component belief-behavior consistency metric. At the population level, it compares an elicited belief ranking of persona-attribute levels against a behavioral ranking of observed mean transfers using Spearman's $\\rho$, and compares elicited versus observed effect sizes using the absolute difference in ANOVA $\\eta^2$. At the individual level, it uses mean absolute error between forecasted and enacted transfer amounts across six Trust Game rounds against fixed Trustee archetypes labeled M1, M3, and M5. The testbed is the Trust Game with an augmented GENAGENTS persona bank that adds Big Five personality traits; single-round simulation isolates the Trustor's decision, while multi-round simulation uses a reasoning-action loop (ReAct). Together these pieces turn \"what the model says\" and \"what the model does\" into concrete numbers that can be compared before a larger simulation is run.","core_discovery":"The central claim is that LLM role-playing agents show systematic belief-behavior inconsistency at both individual and population levels, and that this inconsistency can be measured before running a large simulation. The paper demonstrates this with a Trust Game, where models first state how they think persona attributes such as age, political views, or conscientiousness shape trust and transfer decisions, and then are asked to play the Trustor with those personas. Across three open-weight models, providing full task instructions during belief elicitation did not improve consistency; asking for dollar-amount predictions improved rank-order agreement ($\\rho$) but systematically overestimated effect sizes ($\\eta^2$); feeding a model its own elicited beliefs back as context helped Llama models but essentially failed for Gemma 2 27B; and imposing even weakly perturbed researcher priors reduced alignment across architectures. At the individual level, the mean absolute error between an agent's forecasted and actual transfers grew with the number of rounds forecast ahead. The paper's conclusion is that models may encode plausible beliefs yet fail to apply them consistently, so stated beliefs should not be taken as a guarantee of simulated behavior.","pith_inferences":["A natural extension is to treat belief-behavior consistency as a screening metric in a data-generation pipeline: run a cheap belief-elicitation pass, flag divergent personas, and exclude or re-prompt them before expensive multi-round simulation; the paper demonstrates the diagnostic but does not design this workflow itself.","The Trust Game is a minimal one-dimensional action space; the same framework in richer environments, such as open-ended dialogue or repeated social dilemmas, would show whether inconsistency grows with task complexity, as the near-monotonic horizon effect suggests.","The authors note that reasoning models with extended inference might close the belief-behavior gap; this is directly testable by running their exact prompts on such models.","The prompt-equivalence concern, that belief and role-play formats may activate different task priors, implies a concrete check: counterbalance the two wording styles and see whether measured inconsistency survives."],"forward_implications":["Belief elicitation can serve as an upfront validity check for synthetic human-behavior datasets, flagging roles where stated beliefs and simulated actions are likely to diverge.","Adding task context to belief prompts is not a reliable cure; consistency depends on what is elicited, not just how much information is supplied.","Self-conditioning on a model's own beliefs is model-dependent, so it cannot be assumed to improve alignment in every LLM.","Researcher-imposed priors can actively reduce belief-behavior alignment, which constrains the use of role-playing agents for counterfactual or theory-testing simulations.","Multi-round simulations become less predictable as the forecasting horizon grows, so individual-level forecasts beyond the near term should carry much larger uncertainty."],"supporting_citations":[{"why":"Supplies the Trust Game paradigm that serves as the testbed for all experiments.","marker":"Berg et al., 1995"},{"why":"Introduces generative agents whose persona-based simulation approach the paper extends into an evaluation framework.","marker":"Park et al., 2023"},{"why":"Provides the GENAGENTS persona bank that the paper augments with Big Five personality traits.","marker":"Park et al., 2024"},{"why":"Presents TRUST SIM, which checks stated beliefs against free-form outputs and serves as a comparison point for the paper's focus on enacted behavior.","marker":"Huang et al., 2024"},{"why":"Provides evidence that LLMs can know more than they show, motivating investigation of inconsistency between internal knowledge and behavior.","marker":"Orgad et al., 2024"},{"why":"Uses a simulated Trust Game to reveal name-based biases, establishing the game as a benchmark for LLM trust behavior.","marker":"Wei et al., 2024"},{"why":"Evaluates population-level LLM simulation for policy questions, representing the post-hoc evaluation approach the paper critiques.","marker":"Hou et al., 2025"}],"fun_headline_variants":["LLM agents claim beliefs then betray them in Trust Game","Belief-behavior mismatch in LLM simulations of humans","LLM role-players' beliefs fail to predict their simulated actions","Trust Game exposes gap between LLM agents' words and deeds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the belief-elicitation prompt and the role-play prompt are measuring the same underlying belief; if they instead activate different task priors, the measured inconsistency could be a prompt-format artifact rather than a genuine belief-behavior gap.","fun_headline_variants_meta":{"raw":{"variants":["LLM agents claim beliefs then betray them in Trust Game","Belief-behavior mismatch in LLM simulations of humans","LLM role-players' beliefs fail to predict their simulated actions","Trust Game exposes gap between LLM agents' words and deeds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00089,"raw_usage":{"total_tokens":3914,"prompt_tokens":1097,"completion_tokens":2817,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":713,"completion_tokens_details":{"reasoning_tokens":2747}},"tokens_in":713,"tokens_out":2817,"duration_ms":21741,"temperature":1.0,"reasoning_tokens":2747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:35:19.797412+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reword or counterbalance the two prompt formats: for example, ask the belief-elicitation question in role-play style and the role-play question in belief style, or administer several semantically equivalent variants of the belief prompt. If belief-behavior consistency changes substantially across superficial rewrites, the paper's central gap is at least partly a wording artifact; if it survives, the gap is robust.","supporting_citations":[],"review_version":1}