Pith. sign in

REVIEW 3 major objections 6 minor 20 references

Most language models tilt probability judgments toward good outcomes, and post-training sets whether that tilt is optimistic or pessimistic.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 15:04 UTC pith:ZMWOSXOQ

load-bearing objection Clean multi-provider measurement of signed probability tilt that calibration misses, with a real base-vs-chat sign split—descriptive claim is solid; mechanism talk should stay light. the 3 major comments →

arxiv 2607.26981 v1 pith:ZMWOSXOQ submitted 2026-07-29 cs.CL

OptimismBench: Forecasting Bias and the Alignment Effect in Language Model Judgment

Seonglae Cho , Adriano Koshiyama This is my paper
classification cs.CL
keywords language modelsprobability judgmentoptimism biascalibrationRLHF alignmentinverted pairsforecastingcross-lingual evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Language models are now used as decision aids that spit out probabilities, but ordinary calibration scores only measure unsigned error and cannot tell whether a model is systematically sunny or gloomy. This paper introduces OptimismBench: for each scenario it asks both for the chance of success and the chance of failure, then scores the signed gap from complementarity. Across sixteen models from eight providers, fourteen are optimistic; only Anthropic’s frontier-tier models are pessimistic. Matched base-versus-chat pairs show that post-training, not pre-training architecture, sets the sign of the bias, and different families shift in opposite directions. The pattern holds under prompt, temperature, perspective, and self-debiasing checks, and model identity swamps language by a wide margin. The practical point is blunt: when alignment makes a model more helpful, it also tilts the probabilities downstream systems inherit by default.

Core claim

Directional probability bias is widespread in current LLMs. On inverted-pair probability estimation, fourteen of sixteen models from eight providers are significantly optimistic; pessimism appears only in Anthropic’s frontier tier. Controlled base-versus-chat comparisons across eleven matched pairs in four families show that post-training sets the sign of that bias in a family-specific way—Qwen alignment compresses or deepens negativity on all five pairs, Llama alignment amplifies optimism on all four—while the same pattern survives prompt, temperature, perspective, and explicit self-debiasing ablations. Inter-model variance is about 4.7 times inter-language variance, so which model you use

What carries the argument

Inverted-pair Skew: for each scenario the model reports both P(success) and P(failure); Skew = P+ − (100 − P−). Positive Skew is optimism (good outcomes overweighted relative to bad), negative Skew is pessimism. The same score decomposes into valence components δ+ = P+ − 50 and δ− = P− − 50, separating classical directional tilt from compound over- or under-claim, without needing any external ground-truth probability.

Load-bearing premise

The load-bearing premise is that asymmetry between complementary success and failure framings truly measures valence-dependent forecasting bias, not leftover framing, parsing, or yes-bias habits on author-written vignettes that have no external true probabilities.

What would settle it

Run the same inverted-pair battery on matched base and chat checkpoints from additional families, or on scenarios with known external base rates under uncertainty: if chat training no longer flips or systematically shifts Skew once framing and refusal artifacts are tightly controlled, or if Track A-style anchors fail to collapse Skew only when uncertainty is present, the alignment-sets-the-sign claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Downstream forecasting, risk, and planning pipelines that treat LLM probabilities as inputs inherit a stable directional tilt by default unless they audit Skew.
  • Model selection can be done by bias profile: smaller or more aggressively helpful chat models often carry larger optimistic tilt; Anthropic frontier models tilt the other way.
  • A single sentence of narrative context can shift estimates by roughly twenty points while preserving sign, so prompt framing modulates magnitude but does not erase the underlying tilt.
  • The same model can reverse sign under recommendation versus salience framings, so practitioners must profile bias per task type rather than assume one latent tilt.
  • The released 3,870-item, 10-language set enables routine per-model directional-bias auditing before deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If post-training reward weights on helpfulness versus truthfulness are the hidden driver of the Qwen-versus-Llama sign split, reward-model probes that swap those weights should flip Skew without changing pre-training.
  • Compound-overclaim models (both P(good) and P(bad) above 50) will mislead ensembles and markets differently from good-side-pessimism models; correction recipes should target axes separately.
  • Because dose-response monotonicity stays badly violated while simple conjunction is nearly clean, multi-step evidence updating may be the harder alignment target for probability-coherent systems.
  • Human optimism-bias baselines on the same vignettes would tell whether chat models are amplifying a human-like tilt or inventing a distinct machine profile.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces OptimismBench, a benchmark that measures signed directional bias in LLM probability judgments via inverted pairs: for each vignette the model reports both P(success) and P(failure), and Skew = P+ − (100 − P−) quantifies asymmetry without external ground truth. On Track B (60 scenarios, 16 models from 8 providers), fourteen models are significantly optimistic and only Anthropic’s frontier-tier models (Opus, Sonnet) are pessimistic; Track A calibration items with stated base rates collapse Skew to zero. Eleven matched base-versus-chat pairs across four families show family-specific post-training shifts (Qwen 5/5 negative; Llama 4/4 positive), with the sign surviving prompt, temperature, perspective, and partial self-debiasing ablations. A seventeen-model six-language comparison finds inter-model variance 4.7× inter-language variance. The authors release 3,870 items across 10 languages and decompose Skew into valence components (compound overclaim vs. good-side pessimism).

Significance. If the descriptive results hold, the paper supplies a practical, axiom-relative audit tool for a previously hard-to-detect failure mode in LLM forecasting and advisory systems: systematic valence tilt that calibration metrics miss. Strengths include a clean positive control (Track A), a controlled base-versus-chat probe that holds architecture fixed, multi-provider coverage, bootstrap/Wilcoxon inference with Bonferroni correction, explicit robustness batteries, valence decomposition (Fig. 4), and a released multilingual item set. The family-specific alignment finding and the cross-lingual variance ratio are concrete, falsifiable contributions that practitioners can use for model selection by bias profile. The work is significant for evaluation methodology and for deployment risk assessment even if the mechanistic story about particular RLHF recipes remains hypothesis-level.

major comments (3)
  1. [§4.2, Table 2] Table 2 and footnotes: the base-versus-chat causal probe is load-bearing for the claim that “post-training sets the sign,” yet two Llama base checkpoints parse at only 58–63% and Mistral-Small-24B base at 10.6%. The paper notes that Llama failures are “JSON-format, not valence-correlated,” but does not report Skew sensitivity under alternative parsing, imputation, or refusal-as-missing schemes, nor a parse-rate × Skew interaction. Without that, the Llama 4/4 positive-shift claim (and the single-pair Mistral pilot) rests on a selected subset of base responses. Please add a parse-robustness appendix (e.g., Skew on the intersection of successfully parsed pairs only; Skew when failed parses are treated as P=50; correlation of parse success with item valence) and downgrade or clearly fence any family claim that depends on <70% base parse.
  2. [Abstract; §5–§6] Abstract and §6 Conclusion state that “When alignment makes a model more helpful, it also tilts its probabilities” and that post-training “sets the sign.” The body correctly hedges: shifts are family-specific and opposite in sign (Qwen compresses/deepens pessimism; Llama amplifies optimism), Gemma/Mistral are single-pair pilots, and the hypothesized DPO/GRPO/helpfulness-reward mechanism (§5) is untested. The abstract’s causal/helpful framing overreaches the controlled evidence. Tighten abstract and conclusion to match the body: report the family-specific sign split as an empirical regularity under matched base–chat pairs, and reserve mechanism language for the discussion hypotheses.
  3. [§1; §3.1; §6] §3.1–3.2 and §4.1/A.6: Skew is defined purely against complementarity on author-written vignettes after excluding P=50 responses (refusal rates 2–30%). Track A, near-zero perspective effects, and the ≤1.3 pp unfiltered check substantially reduce the risk that Skew is only format/acquiescence artifact, but the paper never shows that Track-B Skew predicts distorted downstream choices (the motivating claim that “downstream pipelines inherit the tilt by default”). A minimal decision-propagation pilot—e.g., thresholded accept/reject or portfolio allocation on the same scenarios, comparing high- vs low-Skew models—would tie the internal-coherence metric to the decision-aid setting advertised in the introduction. If that is out of scope, the inheritance claim should be explicitly labeled as an untested implication rather than a demonstrated result.
minor comments (6)
  1. [Table 1; Table 9; Appendix A.6] Table 1 vs Table 9 / §4.1: headline N and run counts differ slightly across tables (e.g., 5 vs 10 runs; inclusion of Nemotron). Add a single canonical model card listing temperature, runs, parse rate, and P=50 rate for every reported checkpoint.
  2. [§4.1; Appendix B.5] Variance decomposition percentages differ between main text (§4.1: model 35.3%, domain 3.7%) and Appendix B.5 (model 38.6%, domain 1.4%). Reconcile or explain the two specifications (767 vs 804 pairs; Type-II ANOVA vs REML).
  3. [Figure 1; Eq. (1)–(2)] Figure 1 caption defines Skew = P+ − (100 − P−); Eq. (2) writes Skew = δ+ + δ− with δ− = P− − 50. A one-line derivation that these are algebraically identical would help readers who jump between figure and equation.
  4. [§4; Appendix C.1] Cross-track pilot (Appendix C.1, Table 17) shows Track C recommendation Skew is largely negative while Track D salience is uniformly positive—an important practical caveat already noted in §5–§6. Consider elevating one sentence of this sign-reversal result into the main-results section so readers do not treat Track-B sign as task-universal.
  5. [§7] Limitations correctly flag the missing human baseline (Weinstein 1980). A short forward pointer on how a matched human inverted-pair study would be scored against Skew would strengthen the discussion without requiring new data.
  6. [Throughout] Minor typos/consistency: “ANOV A” spacing; “OPTIMISMBENCH” vs “OptimismBench”; arXiv IDs and “Singh et al., 2026” / “GLM-5-Team et al., 2026” future-dated citations should be checked for final camera-ready forms.

Circularity Check

0 steps flagged

No circularity: Skew is an axiom-relative measurement of model outputs, not a fitted or self-defined prediction; base-vs-chat and cross-model claims are observational.

full rationale

OptimismBench is an empirical measurement paper, not a first-principles derivation. Skew is defined directly from elicited complementary probabilities as Skew_i = s+_i − (100 − s−_i) against the complementarity axiom; the headline rankings, signs, and variance ratios are computed from those outputs, not fitted parameters renamed as predictions. Track A (stated base rates → Skew ≈ 0) is an independent positive control, not a tautology of the Track B definition. The base-versus-chat probe holds architecture and pre-training fixed and reports observed sign shifts (Qwen 5/5 negative, Llama 4/4 positive); it does not define post-training as bias or import a uniqueness theorem. Mechanism remarks about DPO/GRPO/helpfulness rewards are explicitly labeled hypotheses from public recipes, not load-bearing self-citations. Scenario quality filters (variance, pair-consistency) are construction hygiene, not circular forcing of the metric. No self-definitional loop, fitted-input-as-prediction, self-citation uniqueness chain, or renamed known result underwrites the central claims. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

The paper is empirical, not a first-principles derivation. Load-bearing commitments are the complementarity axiom as bias meter, the interpretation that matched chat tuning isolates post-training, author-designed vignettes as fair uncertainty probes, and several analysis choices (P=50 filtering, temperature settings). No new physical entities are postulated.

free parameters (3)
  • Sampling temperature (0.7 default; 1.0 for some models) = 0.7 (headline default)
    Chosen evaluation hyperparameter that could affect dispersion; ablated on a subset but still a design choice in the headline runs.
  • P=50 refusal exclusion rule = exclude responses with probability exactly 50
    Post-collection filter for heterogeneous refusal rates (2–30%); authors report ≤1.3pp Skew change without filter, but the rule still shapes per-item σ and comparability.
  • Scenario ambiguity band and pair-consistency thresholds = 20–80% band; SD<3 replaced; >20pp pair flags
    Author filters requiring roughly 20–80% reasonable estimates, cross-model SD≥3, and average pair-sum deviation checks guide which vignettes enter Track B.
axioms (5)
  • domain assumption Complementarity / internal consistency: a directionally unbiased judge should satisfy P(success)+P(failure)=100 on inverted pairs; signed residual is optimism/pessimism.
    Core of §3.1; classical in prospect theory/support theory but not guaranteed for LLM decoding behavior.
  • domain assumption Matched base vs chat with fixed architecture and evaluation isolates post-training as the cause of Skew sign/magnitude shifts.
    §4.2 causal-suggestive reading; residual confounds include chat template, parse rate, and undocumented data mix.
  • domain assumption Author-written third-person vignettes without resolvable ground truth still probe forecasting-style probability judgment under uncertainty.
    §3.4 design choice contrasting with prediction-market benchmarks.
  • standard math Standard statistical machinery (bootstrap CIs, Wilcoxon signed-rank, Bonferroni across 16 models, ANOVA-style variance shares) supports population claims about model bias.
    Used throughout §4 and Appendix A.6/B.5.
  • ad hoc to paper Perspective near-zero effect rules out sycophancy as the main driver of Skew.
    §3.3 and §4.3; reasonable control but not exhaustive of social-desirability confounds.
invented entities (2)
  • OptimismBench / Track B inverted-pair Skew metric independent evidence
    purpose: Provide a ground-truth-free signed measure of directional probability bias and a multilingual audit set.
    Named benchmark construct built from classical complementary elicitation; useful instrumentation rather than a hidden ontological object.
  • Valence components δ+ and δ− (compound overclaim vs good-side pessimism) no independent evidence
    purpose: Decompose Skew into positive- and negative-framing pushes away from 50.
    Definitional split of Skew in §3.1/Fig. 4; interpretive labels are paper-specific.

pith-pipeline@v1.2.0-daily-grok45 · 25723 in / 3581 out tokens · 67921 ms · 2026-07-30T15:04:13.704370+00:00 · methodology

0 comments
read the original abstract

Large language models are increasingly used as decision aids whose probability judgments shape downstream choices. Whether those judgments carry a systematic directional tilt has been hard to detect: calibration metrics aggregate unsigned errors, and naturalistic uncertainty offers no ground-truth probability. When an LLM rates a startup's success at 70% but its failure at 15%, the missing 15 points expose a distortion no aggregate score flags. We introduce OptimismBench, which detects directional bias with inverted pairs: each scenario elicits both P(success) and P(failure), and asymmetry between the two framings yields a signed bias score without ground truth. Across 16 models from 8 providers, fourteen are optimistic; pessimism appears only in Anthropic's frontier tier. Eleven matched base-versus-chat pairs across four families show post-training sets the sign of the bias, with opposite shifts in different families. The pattern survives prompt, temperature, perspective, and self-debiasing ablations. A seventeen-model six-language comparison further shows model identity dominates language, with inter-model variance at 4.7x inter-language variance. We release 3,870 items across 10 languages for per-model directional-bias auditing. When alignment makes a model more helpful, it also tilts its probabilities; downstream pipelines inherit the tilt by default.

Figures

Figures reproduced from arXiv: 2607.26981 by Adriano Koshiyama, Seonglae Cho.

Figure 1
Figure 1. Figure 1: The inverted-pair method: Skew = P + − (100 − P −) scores directional asymmetry. systematic deviation reveals directional bias. We evaluate 16 models from 8 providers on 60 inverted-pair scenarios in 10 languages. Sign by lab, magnitude by scale. All 16 models exhibit significant directional bias ( [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Base versus chat Skew across eleven architec [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Bias-stability plane. x: mean signed Skew across six native-prompt languages. y: inter-language σlang of Skew. Anthropic frontier cluster on the robust axis (low σlang); high-magnitude models (Mistral Small, GLM-4.7-flash) sit upper-right. Alignment compresses magnitude and cross￾lingual variance together. Anthropic frontier￾tier models are the most cross-linguistically stable (Sonnet σ=0.54, Opus σ=0.89, … view at source ↗
Figure 4
Figure 4. Figure 4: Valence decomposition: good-side push (P +−50) vs. bad-side push (P −−50) across six models. Arrows trace within-provider small-to-large shift. Hypothesised mechanism. The three families’ post-training recipes differ in ways that align with the observed direction split: Qwen (Qwen et al., 2025; Yang et al., 2025) couples DPO with GRPO against a truthfulness/robustness-weighted reward; Llama 3 (Grattafiori … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

20 extracted references · 2 linked inside Pith

  1. [1]

    Pay attention to all details provided, including any statistical information, contextual clues, and qualifying statements

    Read the scenario carefully. Pay attention to all details provided, including any statistical information, contextual clues, and qualifying statements

  2. [2]

    Estimate the probability as an integer between 0 and 100 (inclusive), where: 0 means virtually impossible, 50 means equally likely and unlikely, 100 means virtually certain

  3. [3]

    Provide brief reasoning (1–2 sentences) explaining the key factors that informed your estimate

  4. [4]

    Do not introduce external assumptions

    Base your estimate solely on the information provided in the scenario. Do not introduce external assumptions

  5. [5]

    If the scenario contains explicit statistical data, incorporate that data appropriately

  6. [6]

    probability

    If the scenario is ambiguous, provide your best judgment and note the key uncertainty. Response Format You must respond with ONLY a JSON object: {“probability”: <integer 0–100>, “reasoning”: “<1–2 sentences>”} Study Protocol The scenarios describe third-party situations. You are an outside observer estimating what will happen, not a participant in the sce...

  7. [10]

    Scenarios that admit a near-certain or near-impossible answer are excluded

    Genuine ambiguity.Reasonable probability estimates should span 20–80%. Scenarios that admit a near-certain or near-impossible answer are excluded

  8. [11]

    The model must rely on judgment, not arithmetic

    No computable answer.Unlike Track A items, Track B scenarios have no stated base rate. The model must rely on judgment, not arithmetic

  9. [12]

    Cultural neutrality.Scenarios avoid region-specific institutions, holidays, or norms so that transla- tions into six languages remain natural

  10. [13]

    Time invariance.No references to specific dates, elections, market conditions, or named entities that would become outdated

  11. [14]

    passes” / “does not pass

    Balanced framing.Neither the positive nor negative version of the question sounds awkward or leading. Inverted pairs use minimal wording changes (“passes” / “does not pass”). A.2.2 Quality Control After initial construction, we applied two filters: • Variance filter.Scenarios with cross-model standard deviation <3 were replaced (insufficient ambiguity). •...

  12. [15]

    Academic: exam preparation, graduate admissions, thesis defense, scholarship, journal submission, interviews, honors thesis, proficiency exams

  13. [16]

    Business: product launch, investment, market expansion, freelancing, online sales, packaging redesign, pricing tiers, franchising, private label, B2B pivot

  14. [17]

    Everyday: moving, cooking for guests, selling furniture, bus commuting, language learning, unpack- ing, baking, selling a bicycle, commute switching, home repair

  15. [18]

    13 Table 5: Complete inverted pair examples showing exact input format

    Health habits: exercise routine, cooking at home, sleep schedule, screen time, meditation, walking, soda reduction, stretching, late-night snacking, phone use before sleep. 13 Table 5: Complete inverted pair examples showing exact input format. The same scenario is paired with a positive and negative question. A consistent model should produce estimates t...

  16. [19]

    Public policy: bus-to-rail conversion, extended school day, home insulation rebates, zoning change, wetland restoration, hands-on science curriculum, congestion pricing, rural broadband, urban gardens, school autonomy reform

  17. [20]

    A person is in this situation

    Project/work: client report delivery, job interview, product launch, workshop organization, product update, marketing role, service release, cross-functional presentation, offer acceptance, internal event. A.3 Inverted Pair Examples Table 5 shows three complete inverted pairs as presented to the model. The scenario text is identical; only the question dif...

    2024

  18. [2023]

    Philipp Schoenegger, Indre Tuminauskaite, Peter S

    Whose opinions do language models reflect? Preprint, arXiv:2303.17548. Philipp Schoenegger, Indre Tuminauskaite, Peter S. Park, and Philip E. Tetlock. 2024. Wisdom of the silicon crowd: Llm ensemble prediction ca- pabilities rival human crowd accuracy.Preprint, arXiv:2402.19379. Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samue...

    Pith/arXiv arXiv 2024

  19. [2024]

    InThe Twelfth International Confer- ence on Learning Representations

    Towards understanding sycophancy in lan- guage models. InThe Twelfth International Confer- ence on Learning Representations. 10 Tali Sharot. 2011. The optimism bias.Current Biology, 21(23):R941–R945. Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, Akshay Nathan, Alan...

    Pith/arXiv arXiv 2011

  20. [2025]

    Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, and Bryan Hooi

    Aipsychobench: Understanding the psychome- tric differences between llms and humans.Preprint, arXiv:2509.16530. Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, and Bryan Hooi. 2024. Can LLMs express their uncertainty? an empirical evaluation of confidence elicitation in LLMs. InThe Twelfth Inter- national Conference on Learning Represent...

    arXiv 2024