Paper: OptimismBench: Forecasting Bias and the Alignment Effect in Language Model Judgment Authors: Cho Seonglae, Koshiyama Adriano arXiv: 2607.26981 Link: https://arxiv.org/abs/2607.26981
---
A Strange Arithmetic Problem
Suppose you ask an LLM: "What is the probability this startup succeeds?" It answers 70%.
Then you ask: "What's the probability it fails?" It answers 15%.
70 + 15 = 85. Where did the remaining 15 points go?
These 15 points aren't a rounding error, nor "model uncertainty." They expose a bias that calibration metrics can never catch — directional bias. Calibration metrics only measure "the gap between predicted probability and true frequency," and after positive and negative errors cancel each other out, the tilt becomes invisible. But the "disappearance" of these 15 points precisely shows the model is systematically overestimating positive outcomes.
This paper is built around that observation. The authors propose OptimismBench, which uses a method called "inverted pairs" to measure directional bias in LLM judgment without any ground truth.
---
Method: Inverted Pairs — Bias Detection Without Ground Truth
Core Idea
Traditional bias detection needs "correct answers" — you must know the true success rate to tell whether a model leans optimistic or pessimistic. But many real-world scenarios have no ground truth: whether a startup will succeed, whether a project will ship on time, whether a negotiation will close — the future is unknown.
The cleverness of OptimismBench: it doesn't compare model predictions to real outcomes; it compares the model's answers to two phrasings of the same scenario.
For each scenario, the authors construct two versions:
- Version A: asks for P(success)
- Version B: asks for P(failure)
- Skew > 0 → optimism bias (overestimating positive outcomes)
- Skew < 0 → pessimism bias (overestimating negative outcomes)
- Track A (calibration control): 15 calibration questions with real frequencies, confirming the bias isn't purely a calibration issue
- Track B (probability estimation): the main track, measured with inverted pairs
- Track C (recommendation): asking the model to make recommendations instead of probability judgments, to see whether the bias persists
- Track D (salience): varying the salience of scenario descriptions, to see whether the bias is amplified or suppressed
- 14 models are optimistic — systematically overestimating positive outcomes
- 2 models are pessimistic — both from Anthropic's frontier line (Claude series)
- Post-training determines the sign of the bias — the same base model, after different vendors' post-training, can end up with completely opposite bias directions
- Qwen family: post-training shifted models from pessimistic to optimistic
- Llama family: post-training shifted models from optimistic to pessimistic
- Between-model variance is 4.7x the between-language variance
If the model is "unbiased," P(success) + P(failure) should equal 100% (the complement axiom). If they don't sum to 100%, the gap measures directional bias. The authors define a metric called Skew:
The key of this design: no ground truth needed. The bias comes entirely from the internal consistency of the model's own answers.
Four Tracks
Beyond the main inverted-pair track, the paper designs four tracks to probe the boundary conditions of the bias:
Factor Interventions
The authors also ran four interventions attempting to "dislodge" the bias:
1. Narrative manipulation: changing the scenario's storyline 2. Perspective shifts: having the model answer from different viewpoints 3. Anchoring gradients: providing different anchor values 4. Self-debiasing: letting the model correct its own bias
Result: the bias survived all interventions. Change the prompt, the temperature, the perspective, or ask the model to self-correct — directional bias persists.
---
Findings: 14/16 Optimistic, Only Anthropic's Frontier Line Pessimistic
Main Result
The paper tested 16 models from 8 vendors. Results:
This result is highly counterintuitive. We usually assume "alignment training" makes models more cautious and objective. But the paper finds: alignment training does change bias direction — but different vendors change it in different directions.
The Alignment Gradient
The authors compared 11 "base vs chat" model pairs across 4 model families. Findings:
This shows "optimistic/pessimistic" is not an inherent property of model architecture or pretraining data, but a product shaped during the post-training stage.
Model Identity > Language
The paper also ran a cross-language experiment: 17 models, 6 languages. Result:
Bias Can Flip in Recommendation Scenarios
A particularly notable finding: in Track C (recommendation), the bias sign can flip. A model that is optimistic in probability judgment may become pessimistic in recommendation scenarios.
This means: bias is not an inherent model property, but a joint property of "model × task framing". You can't simply say "Claude is pessimistic" — only "Claude is pessimistic in probability judgment, and possibly optimistic in recommendation scenarios."
---
Why This Matters
1. The Structural Blind Spot of Calibration Metrics
The paper points out an overlooked problem: calibration metrics (e.g., ECE, Brier score) are "unsigned" — they cancel positive and negative errors. A model may be well calibrated overall, yet systematically overestimate positive outcomes and underestimate negative ones; such bias is completely invisible to calibration metrics.
This evokes the "evaluation blind spot law" — you optimize what you measure, and what you don't measure is where problems hide. Calibration metrics measure "average error"; directional bias hides in "the direction of the error."
2. Side Effects of Alignment Training
The paper finds post-training shapes bias direction, differently per family. This suggests RLHF/DPO-style alignment techniques are "shaping model personality" — not just making models safer and more helpful, but making them more optimistic or pessimistic.
This has direct implications for AI safety research: alignment is not neutral. A model trained to be "more helpful" may systematically overestimate success rates in judgment tasks — dangerous for downstream applications like risk assessment, medical decisions, and financial forecasting.
3. Downstream Pipelines Inherit the Bias
The paper's closing line is worth remembering:
> "When alignment makes a model more helpful, it also tilts its probabilities; downstream pipelines inherit the tilt by default."
This means any system using LLMs for probabilistic judgment — from prediction markets to risk assessment to project planning — needs to audit the directional bias of the model it uses. You cannot assume "the model says 70%, so it's 70%"; you need to know whether that 70% has been tilted by an "optimism lean."
---
A Cross-Domain Analogy: Optimism Bias in Human Cognition
"Optimism" in the paper's title is no accident. Human cognitive psychology has a well-known "optimism bias" — people systematically overestimate the probability of positive events and underestimate negative ones. Weinstein's classic 1980 study first documented this systematically.
So LLM optimism bias has an interesting cross-domain isomorphism: models learn human judgment patterns during training, inheriting human cognitive biases along the way. But the paper finds a key difference: humans are universally optimistic, while LLM bias direction is determined by post-training — meaning alignment training could "reverse-calibrate" this bias, but different vendors chose different directions.
This raises a deeper question: should alignment calibrate models to be "unbiased" or "optimistic like humans"? If the goal is competing with humans in prediction markets, "human-like optimism" may actually be more accurate; if the goal is risk assessment, "unbiased" is correct. There's no standard answer, but the paper provides the measurement tools.
---
Limitations
The paper candidly notes several limitations:
1. Family-direction conclusions rest on the Qwen and Llama pairs — Gemma-2-2b and Mistral-Small-24B were single-pair pilots with insufficient statistical power. 2. Uneven cross-language evidence — 6 native-language prompts + 4 English system prompts (DE/FR/HI/JA), the latter being a known confound. 3. Skew measures internal consistency, not agreement with human judgment — an "unbiased" model isn't necessarily an "accurate" model. 4. No human baseline — the paper cites Weinstein (1980) human optimism-bias data but conducted no matched human-subject experiments.
---
My Take
What I admire most about this paper is its methodology: "inverted pairs" make ground-truth-free bias detection possible. It's a classic "solve the problem at a different level" move — traditional bias detection needs ground truth; the authors shifted from "compare against reality" to "compare against itself," bypassing the ground-truth dependency.
But what strikes me more is its practical implication: we are using LLMs as "probabilistic judgment tools" — prediction markets, risk assessment, project planning — yet models carry directional biases that may give all these applications a "family-flavored" tilt. A risk-assessment system built on Claude and one built on Qwen may produce judgments tilted in opposite directions.
This echoes the "evaluation blind spot law": calibration metrics measure "average error"; directional bias hides in "the direction of the error." You measure something, and that's exactly what you stop seeing — a structural blind spot in all evaluation design.
Finally, a question worth pondering: if a model's directional bias is shaped by post-training, then there's tension between "alignment" and "calibration" — alignment makes models more helpful and agreeable, but also tilts their probabilistic judgments. The tilt is a plus for "helpfulness" (users like optimistic answers) and a minus for "accuracy." This tension may have no perfect solution, but at least we can now measure it.
---
Paper link: https://arxiv.org/abs/2607.26981 HTML version: https://arxiv.org/html/2607.26981v1 Dataset: 3,870 test items, 10 languages, open-sourced