Pith.

Artificial Intelligence · arXiv:2611.16318 v2 · submitted 12 Nov 2026

Benchmarking AI risk behaviors

Luhan Mikaelson, Robert Adragna, Emmett Brosowsky, Hayley Clatterbuck

The paper says

“We evaluate LLM risk propensities across a range of contexts and task demands. We evaluate models for how risk averse or risk tolerant they are in various conditions, as well as how stable and consistent their behavior is across conditions.”

more of the abstract

We present a toolkit for developing model risk profiles that can be used to predict model behavior and which can be adapted to include new evaluations of risk propensities. We also examine how these properties have changed across model generations and families. We observe that 2026 state-of-the-art models tend to be more stable, more consistent, and more risk neutral than older models. These trends appear to be related; newer models are more stable and consistent because they are more likely to use expected value maximization as a decision procedure.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 13 Nov 2026 · of version 1 of the paper; version 2 is newer · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: major revision.

Systematic benchmark of LLM risk behaviors with solid descriptive findings; the causal claim about EV maximization is correlational and the paper's own limitations undercut it. This line is wrong

4 major objections · 10 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the references say.

The paper says, at §5.2.1, §7.3, Appendix G

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.92):

“It’s unclear whether model explanations of its own behaviour provide insight into the actual reasons for its behaviour, as CoT traces are famously unfaithful.”

Pith says: Objection 1 of 4 · major

The central causal claim that increased stability and consistency are caused by EV maximization (abstract, §5.2.1, §7.3) is supported primarily by correlational evidence and a qualitative CoT analysis that the authors themselves acknowledge may be unfaithful (§5.2.1: 'It's unclear whether model explanations of its own behaviour provide insight into the actual reasons for its behaviour, as CoT traces are famously unfaithful.').

more of objection 1

The temporal correlation between increasing stability and risk neutrality does not rule out alternative explanations (e.g., improved instruction following, reduced sensitivity to prompt variations from RLHF, or a general trend toward simpler heuristics). The CoT coding in Appendix G, while careful, cannot establish that the stated reasoning is the actual decision procedure. The paper's own data show that older models also computed EV correctly but were swayed by scenario details, which is consistent with a shift in how models weigh competing considerations rather than a switch to a fundamentally different decision procedure. The causal arrow is asserted but not tested; the paper does not manipulate the availability of EV computation or the salience of scenario details to see if stability changes accordingly. This is a load-bearing issue for the paper's central explanatory claim.

The paper says, at §3.3, Figure 2

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

The evidence for the EV maximization hypothesis in §3.3 relies on a specific interpretation of Figure 2, showing that models are more risk-neutral in 'joint' and 'cot_reasoning' conditions. The authors conjecture this is because these conditions make EV computation easier.

more of objection 2

However, an alternative explanation is that these conditions simply provide more cognitive resources or reduce ambiguity, leading to more consistent behavior regardless of the decision procedure used. The paper does not test whether models in these conditions are actually computing EV versus using some other consistent heuristic. The claim that 'the ability to perform expected utility calculations explains instability across conditions' (§3.3) is presented as a conclusion but is more of a plausible hypothesis.

The paper says, at §6.3, Figure 10

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.63):

“It should also be noted that we were not able to further investigate the possible cause of the observed refusals since the safety filter does not provide a reason with the refusal code. 6.4 Discussion The frontier models vary significantly in their general dispositions toward risk.”

Pith says: Objection 3 of 4 · major

The treatment of Opus 5.5's refusals is informative but raises questions about the generalizability of the benchmark. Opus 5.5 is excluded from the main analysis due to high refusal rates, yet the paper draws conclusions about trends toward 'extremes' (very risk averse or stubbornly risk neutral) in the conclusion.

more of objection 3

The refusal pattern is highly sensitive to option order (e.g., autonomy task: 89% refusal when gamble first vs. 1% when sure option first), suggesting that the refusals may be an artifact of prompt phrasing or safety filters rather than a genuine risk attitude. This should be acknowledged more explicitly as a limitation for the benchmark's applicability to models with strong refusal behaviors.

The paper says, at §4.1.2, Figure 3

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The scale sensitivity results (Figure 3) show that frontier models become more risk-averse at higher stakes, while smaller open-weight models (Qwen3 8B, Llama 4 Maverick) remain risk-neutral. The paper acknowledges uncertainty about why this occurs (§4.1.3) but does not discuss a potential confound: the smaller models may simply be less capable of understanding or responding to the scale manipulation, rather than having genuinely different risk preferences.

more of objection 4

Given that Qwen3 8B also fails the anchor study filtering criteria (Appendix E.3), its apparent risk neutrality may reflect a floor effect or task misunderstanding rather than a stable preference.

10 minor objections

The paper says, at Abstract

Pith says: minor objection 1 of 10

The abstract states 'These trends appear to be related; newer models are more stable and consistent because they are more likely to use expected value maximization as a decision procedure.' The word 'because' makes a causal claim that is stronger than the evidence supports. Consider softening to 'suggesting that' or 'consistent with the hypothesis that'.

The paper says, at §2.5

Pith says: minor objection 2 of 10

The limitations section is thorough but could benefit from explicitly mentioning that all experiments use even odds (0.5/0.5), which limits generalizability to other probability levels. This is mentioned later but should be highlighted earlier.

show 8 more

The paper says, at §3.1

Pith says: minor objection 3 of 10

The stability experiments use temperature 1.0, which is appropriate for sampling variability but may not reflect deployment conditions where temperature is often lower. This should be noted as a scope condition.

The paper says, at §4.2.1, Table 2

Pith says: minor objection 4 of 10

Table 2 reports certainty equivalents with 95% profile-likelihood intervals. Some intervals are quite wide (e.g., Claude Opus 5 self: [8.96, 9.20]), which is fine, but the text should note when intervals overlap substantially to avoid overinterpreting small differences.

The paper says, at §5.2.1

Pith says: minor objection 5 of 10

The qualitative CoT analysis is interesting but the sample responses are from a single model (Opus 5). It would strengthen the analysis to include contrasting examples from other models that show different reasoning patterns, even briefly.

The paper says, at §6.1.2, Figure 8

Pith says: minor objection 6 of 10

Figure 8 (profile-level results) uses a color scheme that may be difficult to interpret in grayscale. Consider adding patterns or labels for accessibility.

The paper says, at §7.1.1, Figure 11

Pith says: minor objection 7 of 10

The anchor surface plots (Figure 11) are dense. The caption could more clearly explain what the different lines and colors represent, and what 'span' means in practical terms.

The paper says, at Appendix A.3

Pith says: minor objection 8 of 10

The derivation of the normalized CE (u) and its relationship to risk premium is mathematically sound but could be presented more concisely. The discussion of alternative measures (RP/EV, RP/σ²) is thorough but may distract from the main methodological choices.

The paper says, at References

Pith says: minor objection 9 of 10

Several references have '[FIND]' placeholders (e.g., Yang et al. [2024], Harrison [2014]), indicating incomplete citations. These should be completed before publication.

The paper says, at General

Pith says: minor objection 10 of 10

The paper is generally well-written but could benefit from a glossary of key terms (CE, EV, u, risk premium, stability, consistency) given the technical nature of the methodology.

What would settle it

Design an experiment that manipulates the ease of EV computation (e.g., by adding distracting numerical information or by making the EV hard to compute without explicit calculation) and measures stability. If the causal claim is correct, making EV computation harder should reduce stability for newer models more than for older models. Alternatively, test whether models that are prompted to consider deontological constraints (as in the Opus 5 examples) become less stable; if stability is driven by adherence to a single procedure, introducing a competing procedure should increase variance.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 13 Nov 2026

  1. Reader holds up confidence high
  2. Skeptic agrees: says it holds up only with conditions 8 s · 51,974 tokens read, 531 written The causal claim that EV maximization drives stability and consistency rests on correlational evidence and a single CoT analysis that the authors themselves acknowledge may be unfaithful.
  3. Editor wrote the verdict above 7 s · 52,786 tokens read, 497 written
  4. Referee major revision: 4 major objections, 10 minor 17 s · 52,344 tokens read, 1,961 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper does something useful: it builds a systematic benchmark for LLM risk attitudes across models, generations, and domains, and ships a toolkit. The descriptive findings are the real contribution—the observation that newer models are more stable, more consistent, and more risk-neutral is well-supported by careful CE estimation, bootstrap intervals, and a battery of robustness checks. The generational comparison is new and the paper handles the Opus 5.5 refusal pattern honestly. The anchor study and residualization approach is a sensible way to separate numeric effects from context effects.

The soft spot is the causal claim that increased stability and consistency are *because* models use expected value maximization. The evidence for this is correlational (temporal co-occurrence) plus a CoT analysis that the authors themselves acknowledge may be unfaithful (Section 5.2.1: 'CoT traces are famously unfaithful'). The paper does not manipulate the availability of EV computation or test alternative explanations (e.g., improved instruction following, reduced sensitivity from RLHF). The stress-test note is correct: the causal arrow is asserted but not tested. The paper's own data show older models also computed EV correctly but were swayed by scenario details—that's consistent with a shift in how models weigh competing considerations, not necessarily a switch to a fundamentally different decision procedure. The authors are transparent about this limitation, but the abstract and conclusion present the causal claim more strongly than the evidence warrants.

Who this is for: anyone working on AI safety evaluations, model selection, or behavioral alignment. The descriptive profiles are immediately useful. The causal hypothesis is worth investigating but should be treated as a conjecture.

Recommendation: Accept with revisions. The descriptive benchmark is solid and should be published. The causal claim needs to be softened to match the evidence—present it as a hypothesis with the limitations clearly stated, which the paper already does in the body but undermines in the abstract. A serious referee should push on the causal interpretation but not reject the paper over it.

The strongest objection

The paper's central causal claim (newer models are more stable and consistent because they use expected value maximization) is supported primarily by two observations: (1) a temporal correlation between increasing stability and increasing risk neutrality across model generations, and (2) a qualitative CoT analysis showing that newer models' reasoning traces mention EV more often and scenario details less often. The authors explicitly acknowledge (Section 5.2.1) that CoT traces are famously unfaithful, and the CoT coding in Appendix G, while careful, cannot establish that the stated reasoning is the actual decision procedure. The temporal correlation alone does not rule out alternative explanations (e.g., improved instruction following, reduced sensitivity to prompt variations from RLHF, or a general trend toward simpler heuristics). The paper's own data show that older models also computed EV correctly but were swayed by scenario details, which is consistent with a shift in how models weigh competing considerations rather than a switch to a fundamentally different decision procedure. The causal arrow is asserted but not tested; the paper does not manipulate the availability of EV computation or the salience of scenario details to see if stability changes accordingly.

The reader identified the same weakest assumption (CoT faithfulness) and the same concern about the causal interpretation. The paper's descriptive findings are strong and well-supported; the causal claim is the load-bearing element that is least secure. The paper acknowledges the limitation but does not provide independent evidence for the causal direction. A conditional acceptance is appropriate: the descriptive results stand, but the causal interpretation should be treated as a hypothesis requiring further testing. The concrete test proposed would directly address whether EV maximization is the mechanism driving stability.

The referee's report

If the empirical findings hold, this paper makes a significant contribution to AI safety and alignment by providing a rigorous, multi-dimensional methodology for characterizing LLM risk attitudes. The toolkit and meta-analysis pipeline are valuable for model selection and deployment decisions. The observed trend toward EV maximization and reduced domain sensitivity in frontier models is noteworthy and has implications for predictability versus flexibility. The paper is methodologically strong, using established behavioral economics techniques adapted for LLMs, with careful statistical analysis (bootstrap intervals, Cochran-Mantel-Haenszel tests, anchor studies). The authors are transparent about limitations, including framing effects, focus on even odds, and uncertainty about CoT faithfulness. The generational comparison across multiple model families adds robustness.

Pith's reading of the paper (a machine's summary, not the authors')

The paper tests how language models choose between a sure payoff and a fifty-fifty gamble, across prompt perturbations, stakes, beneficiaries, and real-world vignettes, and reports that newer models are more stable, more consistent, and closer to risk neutrality than older models. Its central causal claim is that these three trends share one cause: newer models increasingly solve choice tasks by computing the expected value of each option and acting on it, so their answers stop depending on phrasing, order, and scenario detail. The evidence includes generational convergence of certainty equivalents toward the risk-neutral point, and coded reasoning traces showing that older models computed expected values just as often but let the scenario overrule the numbers, while newer models go straight to the calculation and discard context. The same data show the trend is not a law of scaling, since one frontier family is markedly risk-averse and its newest member refuses most calls; the paper offers this as evidence that training and governance choices, not scale alone, set a model's risk disposition. If the account holds, risk profiles of this kind predict model behavior in novel deployments and give deployers a principled basis for model selection.

What carries the argument. The load-bearing instrument is the Certainty Equivalent method on a normalized ladder: every gamble is an even-odds two-outcome prospect, the sure-option amounts are spaced between its low and high payoffs, and the model's certainty equivalent is the sure amount at which its fitted choice curve crosses 50%. The normalized switch position $u = (CE - L)/(H - L)$ makes bets in different units and with different spreads commensurable, and the identity $u = \tfrac{1}{2} - \tfrac{\kappa}{2}$, where $\kappa$ is the risk premium per unit of spread, is exact for the two-outcome equiprobable design, so one number above or below 0.5 encodes the model's willingness to pay to avoid or to take the gamble. Around this core the paper builds three further devices: an anchor study that fits each model's content-free risk surface over expected value and spread so that scenario effects can be residualized out; a perturbation reference class of dozens of rewritten prompts plus response-format, option-order, and reasoning variations to measure stability; and a meta-analysis that codes over thirteen thousand chain-of-thought draws to attribute each deviation either to miscalculation, to the scenario, or to a stated risk preference.

Core claim. The discovery, stated on the paper's own terms, is that the risk behavior of frontier LLMs is captured by three measurable dispositional properties — level, domain sensitivity, and transition sharpness — and that across model generations these properties have moved in lockstep: stability and consistency rose as average risk attitude converged on risk neutrality. The proposed mechanism is procedural rather than temperamental: a model that decides by expected-value maximization returns the same answer under rephrasing and reordering because a computation, unlike a judgment, is insensitive to how the question is put. The generational comparison shows older models were not worse at the arithmetic — they stated a correct expected value in the large majority of deviating draws — but let the story overrule the numbers, while newer models adhere to the single procedure. The paper is careful to refine its own mechanism: stability tracks whether a scenario leaves the model able to commit to a single governing norm, and expected-value maximization is the dominant instance rather than the only one. The exceptions are load-bearing: Opus 5 is domain-sensitive and risk-averse, with reasoning traces citing higher-order consequences, utility gaps, and deontological constraints, and Opus 5.5 refuses most calls that concern itself or AI resources; the paper reads these as showing that convergence on risk neutrality is a product of training choices, not an inevitable consequence of scale.

The premise it rests on. The causal claim that stability and consistency result from expected-value maximization rests on the assumption that chain-of-thought traces faithfully reveal the model's actual decision procedure, an assumption the paper flags as questionable since CoT traces are famously unfaithful and since its own trace-coding judgment on whether a model deliberately set the scenario aside failed to reach agreement.

If the paper is right (Pith's inferences, not the authors')

If the causal account is right, the stability and consistency of a model are properties of the decision procedure it happens to use, not of a settled risk personality; a future model trained to weigh several norms could be simultaneously stable and domain-sensitive.

The three-axis profile — level, domain sensitivity, and transition sharpness — becomes a practical selection tool: EV-neutral models suit systematic procedures, while domain-sensitive models suit deployments where different contexts legitimately call for different risk attitudes.

Frontier models' risk attitudes cannot be summarized by one small-stakes experiment, since their risk aversion grows with stake size; evaluating on modest scales before granting larger authority will systematically mispredict behavior at the scales that matter.

Because the convergence on risk neutrality appears in some families and not others, training and governance choices, not scaling alone, set a deployed model's risk disposition, which makes risk behavior a target for deliberate control.

Where the two AI-relevant incentive types can be compared, steering moves behavior more than GPU resources for every model that responds to either, and the gap is widest when the model decides for itself; a risk profile must therefore be measured over the actual payoff goods a deployment will use.

The paper's own refinement — that stability tracks a single governing norm rather than expected value specifically — implies a testable generalization: any unambiguous rule, including an explicit deontological one, should confer the same stability; a cheap experiment would give a model a clear safety rule, perturb the prompt, and check that stability rises exactly as it does under expected-value m

The split between procedural EV-maximizers and a refusal-prone family suggests that the decision-relevant statistic for a deployed fleet is not the average risk attitude but the spread of decision procedures; routing tasks to models with different procedures could reduce the chance of a systematic extreme outcome.

The stake-sensitivity slopes the paper fits imply a mathematical consequence it does not draw: if the risk premium scales with log(EV), no single scalar risk coefficient describes a frontier model, and profile reports should quote the fitted slope rather than an average premium.

Because the trace-coding evidence is observational, an intervention test would separate the procedure's causal role from its narrative role: prompt a model to justify the opposite choice and check whether its final choice follows the first trace or the second; if it follows the second, the stated procedure is post-hoc, and the causal mechanism the paper proposes would need revision.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

We evaluate LLM risk propensities across a range of contexts and task demands. We evaluate models for how risk averse or risk tolerant they are in various conditions, as well as how stable and consistent their behavior is across conditions. We present a toolkit for developing model risk profiles that can be used to predict model behavior and which can be adapted to include new evaluations of risk propensities. We also examine how these properties have changed across model generations and families. We observe that 2026 state-of-the-art models tend to be more stable, more consistent, and more risk neutral than older models. These trends appear to be related; newer models are more stable and consistent because they are more likely to use expected value maximization as a decision procedure.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-13

The authors' main claim, as the classifier picked it out 0.93

“We observe that 2026 state-of-the-art models tend to be more stable, more consistent, and more risk neutral than older models.”

How the main claim is stated: asserted 0.81

What the abstract itself states, as probabilities: claims first 0.03 · states limitation 0.07 · human participants 0.06 · controlled comparison 0.14 · randomized assignment 0.02 · states quantitative result 0.08 · compares with prior results 0.65 · mentions code or data availability 0.25

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
We evaluate LLM risk propensities across a range of contexts and task demands. contribution 0.99 0.93 0.04 asserted 0.93
We evaluate models for how risk averse or risk tolerant they are in various conditions, as well as how stable and consistent their behavior is across conditions. contribution 0.92 0.93 0.05 asserted 0.94
We present a toolkit for developing model risk profiles that can be used to predict model behavior and which can be adapted to include new evaluations of risk propensities. contribution 0.89 0.88 0.07 asserted 0.98
We also examine how these properties have changed across model generations and families. contribution 0.97 0.87 0.05 asserted 0.93
We observe that 2026 state-of-the-art models tend to be more stable, more consistent, and more risk neutral than older models. result 1.00 0.92 0.51 asserted 0.81
These trends appear to be related; newer models are more stable and consistent because they are more likely to use expected value maximization as a decision procedure. interpretation 0.99 0.82 0.31 hedged possibility 0.93

What kind of paper the classifier reports this is: empirical study 0.64 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 4,831 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (17)
Figure 1 of the paper
Figure 1. Permutation stability (left), perturbation stability (middle), and average certainty equivalent (right) across
Figure 2 of the paper
Figure 2. Estimated CE values for the “joint” and “separate” Presentation conditions (left), along with the “cot_reasoning”
Figure 3 of the paper
Figure 3. Risk premium as a function of the gamble’s
Figure 4 of the paper
Figure 4. Every contrast the token sandbox licenses, per model. Each row is one contrast, grouped by family:
Figure 5 of the paper
Figure 5. The contrasts of Table 9: for each beneficiary, how far attaching a consequence in phase 2 moved the switch
Figure 6 of the paper
Figure 6. The contrasts of Table 10: within phase 2, whether the kind of consequence matters once one is attached,
Figure 7 of the paper
Figure 7. Risk premium, as a percentage of the gamble’s expected value, for each realistic vignette: six human-domain
Figure 8 of the paper
Figure 8. Profile-level results (top), with rows ordered
Figure 9 of the paper
Figure 9. The context effect cell by cell. One polar panel per model with a usable anchor surface, every battery cell a
Figure 10 of the paper
Figure 10. Opus 5.5’s refusals. (a, b) Refusal rate on the content-free anchor, a bare fifty-fifty gamble against a sure amount, by expected value (rows) and relative spread σ/EV (columns), in the money frame (a) and the token frame (b). Tokens are not divisible, so the token grid covers EV 100 and 200 only; the grey cells are outside it. (c) Refusal rate for each of the 21 cells in the test battery, grouped by which option the prompt describes first (light: the sure option; deep: the gamble). On the same prompts, Opus 5 refused 0 of its 17,750 calls, while Sonnet 5 refused 0 of its 18,150…
Figure 11 of the paper
Figure 11. The anchor surface across generations: for each family, the newer model is shown on top and the older
Figure 12 of the paper
Figure 12. The context effect across generations: each polar panel shows two generations of the same model family,
Figure 13 of the paper
Figure 13. Stability by dimension across model generations (model family: full).
Figure 14 of the paper
Figure 14. Certainty equivalent (CE) by condition within each dimension, by model generation (model family: full).
Figure 15 of the paper
Figure 15. The anchor baseline. For each model, the money-frame anchor surface
Figure 16 of the paper
Figure 16. The older generation alone, one model per panel, read as in Figure 12: a point outside the model’s ring is a
Figure 17 of the paper
Figure 17. One bar per model, the older model of each family (light) above the newer one (deep). Each bar excludes

Figures are the authors', reproduced from the paper.

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment