acceptodds
Withdrawn from ICLR 2027

Sensitivity or Criterion? Measuring and Moving When Language Models Act

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Before an agent acts it must decide whether any action is warranted. Benchmarks that ask this question report one accuracy, which cannot distinguish a model that fails to recognize a warranted action from one that recognizes it but is reluctant to act. Signal-detection theory separates the two as sensitivity and criterion, but a criterion estimated from authored binary labels depends on each benchmark's mix of examples and cannot be compared across benchmarks. We introduce FlipAct, a diagnostic built on a qualitative representation of what tools do to physical states. A rule over annotated tool properties computes each question's answer and a graded evidence score, with the Yes/No boundary fixed at zero for every model, so every answer is computed, not written by a model. Questions are built from 2x2 quads that hold the required knowledge fixed while the state flips. The slope of a model's Yes-rate against the score is its sensitivity, and the score where it crosses one half is its criterion. Across 44 open models, sensitivity rises with size within a family, and its ranking agrees with three of four external when-to-act benchmarks. Models respond to the sign of the evidence rather than its size, accepting actions that leave the state no better about as often as helpful ones. When Yes means act, most models are conservative and stay so when the question is rephrased. In some models this is a preference for answering No, and the criterion is not set by size. Activation steering moves the criterion with the weights fixed, and distilling a model from its own steered copy moves it in the weights, without a significant change in sensitivity. The distilled shift carries to six external benchmarks with sensitivity kept, including two without a Yes/No answer. Accuracy rises where the model started conservative and falls where it started permissive.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_dzqdOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

If you want to benchmark a model in tool use, you will roughly want to split it into: if the model can tell when it has to use a tool vs if the model is too reluctant or eager to use it. The paper builds a test set for this where each question asks whether a tool is a wise choice in a physical situation (such as a humidifier in a very dry or a very wet room), with the right answer worked out by some fixed rule. Across 44 open models, larger models within a family are better at telling the cases apart, and most models are too reluctant to say yes. The authors then push models toward acting, by steering their activations and by training each model to copy its steered version, which makes them act more on other tool-use benchmarks without making them any better at telling the cases apart.

Critical strengths
  • It seems like a good idea to have a criterion comparable across benchmarks, and to try to use convergent evidence such as from steering or graded evidence.
Critical weaknesses
  • One big piece of feedback is that a main selling point is a criterion comparable across benchmarks, but it doesn’t carry over to other benchmarks: FLIPACT’s criterion is uncorrelated with the criteria measured on When2Call, QuestBench and When2Tool. The authors reframe it as “a stable reading of each model on FLIPACT rather than a trait,” but this importantly concedes the comparability that motivated the fixed scale!
  • I found the prose very confusing, and hard to parse. The abstract, for example, reads as entirely or almost entirely AI generated.
Questions for authors

NA

Reviewer_hBnrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper looks at two reasons a model can get "should I act?" decisions wrong. It may not be able to tell when an action would help and when it would hurt (low sensitivity), or it may tell the difference but still be too cautious or too eager to act (its criterion). To measure these separately, the authors build FlipAct, a set of 1,880 Yes/No questions such as "the room is very dry, a person uses a humidifier, is this a wise choice?". Each tool is asked about in a situation it fixes and in one it makes worse, so the model has to read the situation. The right answers come from a fixed rule, which also gives each question a score from -3 (clearly a bad idea) to +3 (clearly a good idea). How sharply a model's answers change from No to Yes along this score is its sensitivity, and how much evidence it needs before saying Yes is its criterion.

Across 44 open models, bigger models are more sensitive, and sensitivity on FlipAct agrees with three of four other benchmarks. Most models are cautious, and for some (the Qwen3.5 models and GLM-4-32B) this is mostly a preference for answering "No". The authors then make models more willing to act without changing their sensitivity, by steering their internal activations and by training each model to imitate its steered version, and this carries over to 6 other benchmarks.

Critical strengths
  1. The questions are designed so that a model has to read the situation. Because every tool is asked about both in a situation it fixes and in one it makes worse, knowing what the tool is for isn't enough to answer correctly. The right answers come from a rule rather than being written by a model, and a simple classifier that only sees the wording of the questions can't guess them. The scores are also consistent, since two random halves of the questions give almost the same result for each model, and they hold up when the questions are reworded. This matters because it makes the sensitivity measure hard to game, and the same approach can be extended to new kinds of situations cheaply.

  2. Asking each question the other way round separates real caution from a preference for saying "No". The authors also ask "would using it be unwise?", where answering No means the model would act. A genuinely cautious model should stay cautious under both wordings, while one that just prefers the word "No" should flip. This shows that the caution of the Qwen3.5 models and GLM-4-32B is mostly a preference for "No", while Gemma, phi-4 and Mistral stay cautious either way. I think this is the clearest new insight in the paper, and the same check could be used on any Yes/No benchmark.

Critical weaknesses
  1. Nobody checks whether people agree with the rule's answers, so it isn't clear what the criterion measures. The rule only asks whether a tool moves the situation towards normal, and ignores safety and side effects. The best model only gets 69% of questions right, and the typical model says Yes to only about half of the actions the rule calls wise. Without a human baseline I can't tell whether the models are too cautious or the rule calls some questionable actions wise. For 9 of the 21 models the share of Yes answers never reaches 50% at any score, so their criterion comes from extending the fitted curve beyond the data. A model's criterion on FlipAct also doesn't predict its criterion on 3/4 other benchmarks, even though comparing models across benchmarks was the main reason for building FlipAct.

  2. Making models "more willing to act" mostly just makes them say Yes more. The steering direction is learned from the model's own Yes and No answers, so it captures "about to say Yes" rather than "acting is warranted". When the questions are reversed so that No means acting, most of the change in the trained models reverses too. A model trained only to say Yes more shifts the other Yes/No benchmarks just as much or more, and adding a fixed boost to the "Yes" answer works as well as steering in 19 of 20 cases. The effect beyond saying Yes is small and only shown for two models on two benchmarks.

Questions for authors
  1. Could you ask people to answer a sample of FlipAct questions (for example 200, spread across the scores and question types) and report how often they say Yes at each score? If people agree with the rule, the finding that models are cautious stands. If people are also cautious, especially on questions the rule says Yes to, the caution comes from the benchmark rather than the models.
  2. What do the questions look like that the rule says Yes to but most models reject? Breaking these down by where the tools came from (for example medical and chemical word lists compared with household objects), with examples, would show whether the models' refusals are reasonable.
  3. Since a model's criterion on FlipAct doesn't predict its criterion on other benchmarks, what evidence is there that it means something outside FlipAct?
  4. How much of the agreement between FlipAct sensitivity and the other benchmarks just reflects how capable a model is in general? Once a general knowledge score (MMLU) is taken into account, the correlation drops to about 0.53 and is only borderline significant across 15 models. More models would make this clearer.
  5. The paper says the dataset and code "will be released". Can they be made available during the review period?
Reviewer_C3mrOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

FlipAct asks yes/no questions about whether to use a tool in a given physical state. The questions come in 2×2 quads that hold the required knowledge fixed while the state flips. A rule over LLM-annotated tool properties computes each answer and a graded evidence score ΔU, with the Yes/No boundary fixed at zero. The authors fit the Yes-rate as a logistic in ΔU. They read the slope as sensitivity and the crossing point τ as criterion. Across 44 open models they report that sensitivity rises with size and tracks external d′ on three of four benchmarks, that models respond to the sign of ΔU rather than its size, and that most models are “conservative”. They also report that activation steering, and distillation from a steered copy, shift τ with sensitivity roughly preserved, and that the shift transfers to six external benchmarks. The paper reads as primarily AI-written slop and I recommend desk-rejection on this basis alone.

Critical strengths

The paper attempts many benchmarks and coverage is wide: 44 open models, four external when-to-act benchmarks for the sensitivity comparison, and six external benchmarks for the transfer of the distilled shift. The authors also seem to honestly include and note failure modes.

Critical weaknesses

The paper reads as primarily AI-written slop and I recommend desk-rejection on this basis alone. Beyond that impression, I have concrete concerns about the central measurement and presentation. Everything rests on LLM-annotated tool properties, validated by re-annotating only 150 rows (67.3% same magnitude band). All stimuli come from one template in one physical domain, so I would be cautious about reading the results as a general statement about when language models act and we thus don’t learn much.

On presentation, I found the paper very hard to follow from Section 4 onward: the main text depends on dozens of pointers to a long appendix for the definitions of the quantities it reports (for example which models count as having a “well-determined criterion”), and the argument is difficult to reconstruct without it.

Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.