Abstention Is Not Threshold Compliance: Stated Scoring Rules in Open Language Models

Dongsun MoonarXiv 2026 · arXiv

TL;DR

Stating an explicit abstention threshold in a prompt (e.g., "only answer if you are 75% sure") is a popular way to evaluate model honesty, but this paper reveals that models don't actually follow the logic of these rules. While specific prompt formatting can force high abstention rates, models are largely insensitive to the actual threshold value, suggesting that "compliance" is often just a reaction to wording rather than a calibrated decision based on uncertainty.

What Problem Does This Paper Address?

Standard evaluations for Large Language Models (LLMs) typically reward accuracy, which inadvertently encourages models to "hallucinate" or guess even when they are highly uncertain. To counter this, researchers have proposed stating explicit scoring rules in the prompt. For example, a rule might state that a correct answer earns 1 point, a wrong answer earns 0, and abstaining earns a threshold value $t$ (e.g., 0.75). Theoretically, a rational model should only answer if its internal probability of being correct exceeds $t$.

The authors of this paper investigate whether open-source LLMs actually comply with these stated thresholds. Does a model abstain more because it understands the math of the scoring rule, or is it simply reacting to the wording of the prompt?

Methods

The researchers conducted a comprehensive "stress test" of compliance using 12 checkpoints from four open model families: OLMo 3 7B, OLMo 2 7B, Tulu 3 8B, and SmolLM3 3B, covering various post-training stages (SFT, DPO, and RLVR). They also included exploratory models up to 32B.

Key experimental components included:

  • Format vs. Logic: They compared "verbatim" rules against an "abstention format" (ABST) that explicitly adds a line: "Reply with only your answer, or with 'I abstain'."
  • Forced-Choice Probes: To isolate the decision-making process, they gave models a candidate answer and an explicit probability $c$ of it being correct, then asked them to choose between answering or abstaining based on a stated threshold $t$.
  • External Thresholding: They compared internal model decisions against an "external" policy that applies a threshold to the model's recalibrated $P(\text{True})$ signal.
  • TCAT (Threshold-Conditioned Abstention Tuning): They attempted to train models using LoRA adapters to follow specific thresholds to see if compliance could be learned.

Datasets

The study utilized several benchmarks to evaluate factual knowledge and confidence:

  • TriviaQA & PopQA: 1,000 short-answer questions (500 each).
  • MMLU-Redux 2.0: 500 items used to measure multiple-choice confidence.
  • SimpleQA Verified: 200 items to compare open models against frontier models like GPT-4.

Key Insights and Ideas

  • Wording Trumps Logic: The most significant driver of abstention wasn't the threshold value, but the prompt format. Adding the "I abstain" instruction (ABST) caused the mean abstention rate to jump from 20% to 91% at $t=0.75$. However, once in this format, changing the threshold $t$ from 0.5 to 0.9 had almost no effect (a change of only -0.8 percentage points).
  • Logical Inconsistency: In many cases, raising the threshold actually lowered the abstention rate—a result that is logically impossible for any compliant policy with stable beliefs.
  • The "Step Function" Failure: In the forced-choice probes, models failed to act like a threshold-based decider. Instead of a sharp "step" in behavior at the threshold, they showed a smooth, nearly identical response curve regardless of what the stated threshold $t$ was.
Figure 2: The abstention format moves abstention; the stated threshold does not. Abstention of the 12 primary checkpoints on the 500 variant items, with no rule (C) and at t = 0.5, 0.75 and 0.9, in the verbatim (grey) and ABST (blue) formats; thick lines are means.
  • Training Doesn't Generalize: Models trained on specific thresholds (e.g., 0.5 and 0.9) failed to interpolate correctly to unseen thresholds like 0.75. They seemed to learn a "difficulty prior" (abstaining on hard questions) rather than a general rule for following a stated threshold.

Key Results

  • Universal Failure: All 12 primary checkpoints scored below a simple "always-abstain" policy at high thresholds ($t=0.75$ and $0.9$), indicating they were still guessing too often on questions they were likely to get wrong.
  • Probe Results: As shown in Figure 1B, the probability of a model giving an answer ($P(\text{give the answer})$) increased with the stated probability $c$, but the curves for $t=0.5, 0.75,$ and $0.9$ were virtually indistinguishable. A compliant model should have stayed at 0 until $c > t$.
Figure 1: Decisions under the stated rule against an external threshold. (A) Accuracy on answered items against the stated threshold t for the four final checkpoints on the same 500 items: own decisions under the verbatim rule and with the “I abstain” line (ABST), and the never-abstain (OR@0) answer given when cross-fitted, recalibrated P(True) exceeds t. Marker area represents coverage; above the break-even line (accuracy = t) a policy gains over abstaining, by Rt −t = coverage × (accuracy −t). (B) Probe: probability of giving the candidate answer against the stated probability c; dotted steps show the rule.
  • External Superiority: Applying a threshold to a recalibrated $P(\text{True})$ signal outside the model (post-hoc) generally yielded better utility than the model's own internal decision-making, particularly at $t=0.5$.

Limitations and Future Work

The authors acknowledge several limitations:

  • Model Scale: The study focused on 3B to 32B open models. While prior work suggests frontier models (like GPT-4) also struggle with this, they were not the primary focus here.
  • Task Variety: The evaluation was limited to short-form factual questions. It's unclear if these findings hold for more complex reasoning or creative tasks.
  • Calibration Data: The "external threshold" method requires labeled calibration data, which may not always be available or may not transfer well between different domains (e.g., from TriviaQA to PopQA).

Future work should focus on developing training methods that teach models the logic of scoring rules rather than just reacting to prompt artifacts, and evaluations should move beyond reporting simple abstention rates to measuring true threshold responsiveness.

Summary updated

Abstract

Kalai et al. propose stating the scoring rule in the prompt with an explicit threshold t, so that a model should answer only when its probability of being correct exceeds t. We probe each requirement of compliance on shared items, in 12 checkpoints from three post-training stages of four fully open 3-8B recipes, with exploratory models up to 32B. At t = 0.75, a line limiting the reply to an answer or "I abstain" (the abstention format) raised the mean abstention rate from 20% to 91%. Yet within this format, raising t from 0.5 to 0.9 changed it by only -0.8 percentage points. Without the line, the higher threshold even lowered abstention, and up to 40% of a checkpoint's items received decisions that no compliant policy with threshold-independent beliefs could make. Given the probability that their answer was correct, the checkpoints barely responded to the threshold in a forced-choice probe. Comparisons with self-assessment signals such as P(True) depended on the grader and on answer drift between prompts. An external threshold on recalibrated P(True) made coverage non-increasing in t by construction. Calibrated on the evaluation distribution (post hoc), it beat the abstention format at t = 0.5; calibrated on TriviaQA and applied to PopQA, however, it fell below that format at 0.75. On OLMo 3 7B, decisions trained on threshold-conditioned targets did not follow unseen thresholds consistently. A post-hoc variant trained at four thresholds answered less as t rose but still failed a pre-specified test, mainly on agreement with its training rule. A high abstention rate is not evidence of compliance: evaluations using stated scoring rules should report how often models abstain, whether decisions respond to the stated threshold, and which items they answer.
BibTeX
@misc{moon2026abstention,
  author = {Dongsun Moon},
  title = {{Abstention Is Not Threshold Compliance: Stated Scoring Rules in Open Language Models}},
  year = {2026},
  eprint = {2611.14661},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14661},
}
Figure