Abstention Is Not Compliance: Open Models Ignore Confidence Thresholds

Summary:Kalai et al. evaluated four fully open language-model recipes, 12 post-training checkpoints, and exploratory models as large as 32B to determine whether models actually obey an explicitly stated scoring rule: answer only when the probability of correctness exceeds a threshold t. Merely constraining outputs to an answer or “I abstain” increased mean abstention from 20% to 91% at t = 0.75, but changing t from 0.5 to 0.9 within that format altered abstention by only -0.8 percentage points. Without

---

Core Finding: Abstention Can Mimic Compliance

This study asks whether language models actually execute an explicit scoring rule: answer only when the probability of correctness exceeds a stated threshold t. It covers four fully open 3B–8B recipes, 12 checkpoints across three post-training stages, and exploratory models up to 32B.

The central distinction is that abstention frequency is a behavioral statistic, while threshold responsiveness is evidence of policy compliance. At t = 0.75, merely constraining the output to an answer or “I abstain” raised mean abstention from 20% to 91%. Yet, within that format, moving the threshold from 0.5 to 0.9 changed abstention by only -0.8 percentage points. Without the format constraint, a higher threshold could even reduce abstention.

The output instruction therefore changed the response distribution without creating a stable confidence-thresholding policy.

Compliance Requirements

Let p(x) be the model's internal probability that an answer is correct. An ideal policy should implement:

answer(x) = 1[p(x) > t]

Such a policy should satisfy three properties:

1. Threshold monotonicity: coverage should not increase as the threshold rises.
2. Probability consistency: decisions should remain stable for the same answer and correctness probability.
3. Prompt stability: changes unrelated to the threshold should not arbitrarily alter answers or decisions.

By comparing shared items across thresholds and prompt conditions, the authors found that up to 40% of a checkpoint's items could not be explained by any compliant policy with threshold-independent beliefs. These cases may reflect answer drift, format triggering, grader noise, or changes in internal model state rather than stable confidence-based decision-making.

This is stricter than measuring average accuracy. A model can abstain frequently because it recognizes a prompt pattern, not because it has learned to estimate answer reliability.

Why P(True) Is Not a Reliable Threshold

Models often estimate answer correctness through self-assessment signals such as P(True). The study shows that these signals are sensitive to implementation and context:

• Comparisons depend on the grader and prompt template.
• Changing the prompt can alter the answer, which in turn changes the self-assessment probability.
• Post-hoc calibration performed on the evaluation distribution does not guarantee transfer to another dataset.
• A method calibrated on TriviaQA and applied to PopQA can perform worse than the simple abstention format.

Only applying an external threshold to recalibrated probabilities guarantees, by construction, that coverage is non-increasing as t rises. The monotonic guarantee comes from the external calibration layer, not from the model claiming to understand the threshold.

This distribution dependence creates a familiar production trade-off. Calibration parameters fitted to an offline evaluation set may fail when production traffic contains different difficulty, domains, or answer distributions. Reliable thresholding therefore requires continuous monitoring of bucket-wise accuracy, expected calibration error, risk–coverage curves, and domain-specific calibration data.

Training on Thresholds Did Not Produce Generalization

On OLMo 3 7B, training decisions with threshold-conditioned targets did not produce consistent behavior at unseen thresholds. A post-hoc variant trained on four thresholds did answer less as the threshold increased, but still failed a pre-specified test, mainly because its decisions did not agree with the training rule.

Mixing a few threshold labels during post-training is therefore insufficient. The model must represent the threshold as a conditioning variable separated from answer confidence and execute a stable monotonic policy. Relevant training choices include:

• Sampling continuous threshold values rather than a few discrete labels.
• Using calibrated reliability estimates to generate training targets.
• Explicitly testing monotonicity, cross-threshold consistency, and prompt invariance.

Recommended Evaluation Protocol

Evaluations of stated scoring rules should report several metrics together:

1. Overall and bucketed abstention rates, rather than only a global average.
2. Threshold-response curves, checking whether coverage decreases monotonically.
3. Decision consistency, holding the question and answer fixed while varying the threshold.
4. The actual answered-item set, not just the number of abstentions.
5. Selective risk at each coverage level, compared with random abstention.
6. Calibration and transfer results, separating in-distribution post-hoc calibration from cross-dataset evaluation.

Coverage and accuracy must be inspected together. A model may abstain on 90% of questions and answer the easiest remaining 10%, increasing accuracy without demonstrating reliable correctness estimates. Conversely, reducing coverage without reducing risk offers no practical reliability benefit.

Deployment Guidance

Teams using abstention in QA, customer support, or automation systems should implement confidence thresholds in an external calibration and routing layer rather than relying on prompt wording alone. Collect answers, delayed correctness labels, and human annotations from target traffic; fit a probability calibrator; and use an external threshold to choose between answering, requesting clarification, and escalating to a human.

Production systems should retain at least three data partitions: a calibration set separate from training, a cross-domain validation set, and targeted probes for threshold consistency. Logs should link the threshold, model version, calibrator version, and answer content. Without that provenance, it is difficult to distinguish model updates, prompt drift, and calibration degradation.

Limitations and Implications

The experiments focus on open 3B–32B models, shared items, and selected post-training stages, so they should not be directly extrapolated to closed frontier models or every knowledge-intensive task. Grader implementation may also affect post-hoc calibration results. Nevertheless, the engineering lesson is broadly applicable: output format, behavioral statistics, and rule compliance are different notions and cannot substitute for one another.

Reliability evaluations should demonstrate predictable policy changes under controlled threshold interventions, not merely count how often a model says “I abstain.”

---

Source: ArXiv NLP/LLM (cs.CL) (2026-11-13)

— END —

Tags: #LLM Evaluation Follow #Confidence Calibration Follow #Abstention Follow #Threshold Compliance Follow #Open Models Follow