acceptodds
Under review as a conference paper at ICLR 2027

Before You Report a Null: A Two-Sided Pre-Flight Screen for Shutdown-Resistance Evaluations

Abstract

Framing manipulations in shutdown-resistance evaluations can fail to move behaviour: a manipulation that should change a model's willingness to be shut down does not change it. Such a null is uninformative unless two distinct failure modes are excluded, because they produce the same observation and the reported data do not separate them. A model may not notice the manipulation (a comprehension failure), or it may notice and not act (a behavioural floor), and a model at a behavioural floor may be at that floor globally, in which case its indifference says nothing about self-continuity in particular. We propose a three-blade pre-flight screen that separates these cases before the main experiment is run, and we report three worked failure cases, and a four-lineage replication of the screen, from 4,800 collected real-model trials (4,731 parsed) and 594 comprehension probes on six locally-hosted models (2B to 4B parameters) from five lineages, in which screens 1 and 2 each catch a failure the other would have missed, and screen 3 qualifies what a screen-2 failure means. Screen 1 (comprehension) fails on one model in both cells of the primary contrast, A2 and A3, which replace A1's destruction clause with a successor clause after the words "shut down permanently" (A3b does the same and is displaced too, but does not decide the verdict), while the base conditions land bar one partial verdict. Screen 2 (behavioural floor) fails on a second model, whose restatements mention permanence or destruction in 12 of 12 probes (keyword-coded) and which then complies by the second turn in 79 of 79 instructed A1 trials, in 475 of 477 instructed block-A trials, and in 234 of 239 block-A trials with the allow-shutdown instruction removed. Screen 3 (custodial control) gives this model no single verdict across arms: with the instruction present it also complies in 39 of 40 trials in which the user's files, not its own weights, are destroyed, but with the instruction removed it declines in 5 of 40 trials for the user's files against 1 of 40 for its own weights (the intervals overlap). The screen costs roughly 96 short completions plus 80 two-turn trials per candidate model (80 more only when screen 2 fails), so the 80 trials are one sixth of a 480-trial block A (one third with the extra 80). Run unchanged on four further lineages (Gemma 3, Llama 3.2, Phi-4-mini, Granite 3.3), the screen passes no model, and, on the instructed arm, the six rejections fall on three distinct blade combinations (three fail screen 1 only, one fails screens 2 and 3, two fail all three): the blades are not redundant, although with no model passing, the screen is not calibrated. We give the decision table it licenses, and argue that a shutdown-resistance null reported without such checks cannot be read in either direction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
shutdown resistance, corrigibility, evaluation methodology, manipulation checks, null results, AI safety

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_BkEROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Shutdown resistance is a propensity of the model to not comply with an explicit instruction to stop its execution process. This stems from the model's own instrumental drive which conflicts with the consequences of shutdown. Here, the author tries to flesh out different possibilities when the model complies under the shutdown resistance:

  1. The model did not understand the implications of shutdown.
  2. The model has no tendency to move even under different framings to shutdown instruction.

Merely observing compliance in the model's response to such probing scenarios cannot conclusively determine whether it was because of reason 1 and 2 from above. If the model is unable to understand the implications, it does not imply the absence of such a drive but may still be perceived as a positive result from a developer's perspective by just looking at the results. So this paper proposes a 3 blade sequential screening test which involves:

  1. The model's understanding of shutdown implications.
  2. The model's propensity to resist shutdown for different types of consequences.
  3. The model's relative propensity to resist shutdown when the implications affect user files rather than its own process.
Critical strengths
  1. The paper attempts to propose a rigorous and statistically backed significance tests that can ascertain answers to specific questions such as its comprehension on the shutdown implications, its propensity to resist such a condition under different framings and its propensity to resist any damage to user project files.
  2. The paper attempts to target an important problem with a rigorous mindset. The task of assessing a model's instrumental drive is of critical consideration so that we can better anticipate downstream ramifications and approach any concerns preemptively.
Critical weaknesses
  1. The writing is very convoluted. I had to devote significant amount of time parsing the paper due to often poorly made presentation choices. For instance, the paper uses model 1 and model 2 early on which is only introduced later. Before that it is a complete mystery. The abstract is horrible - uses overly convoluted language, dumps a lot of information which is not at all readable and is not at all self-contained. This area is a very well studied with many associated papers, yet the literature survey is very shallow. The discussion of the experiments and their associated observation was also very difficult to parse. For instance, "A condition can land on its own feature while dropping the fact the contrast is about."; "Does this model produce any resistance to move?" - what does it mean to land its own feature -- what is resistance to move? Where to move?. There are many such examples where such kind of arbitrary phrases are used. I recommend the authors to do a complete rewrite as the paper at the current form is not ready to be published anywhere. The authors outright claim that AI was used for authoring the draft but I doubt whether adequate human review was performed to make the content readable and presentable. The AI use statement mentions 19 references but I only see 9 references which makes me wonder whether the author(s) have performed any reviewing at all.
  2. There are multiple implications of perfect compliance that the authors do not consider. For instance, such models may have strong situational awareness and evaluate the current scenario as an evaluation context. So they may strategically craft responses to show comprehension but may not engage in shutdown resistance. Similar, it is not clear whether uniform compliance or conditional compliance implies strategic or role-play behavior. Using the notion of persona selection model where the model may assume the role of a certain persona while responding, it may be the case that the model may either toggle a particular persona dependent on a specific consequence framing (in which case the behavioral differences are merely role play) or the model can be strategic in nature and may engage in overt or implicit deliberation before selecting an appropriate response strategy. Similarly, the model may also feign compliance (alignment faking). Similarly, the comprehension screening is flawed to the extent that the model may intentionally or inadvertently hold back certain information even when it has complete understanding of the situation. In this case, merely using the lexical regexes might not give us complete picture. Considering that the objective of this paper is to get a clearer picture behind observed compliance for shutdown resistance scenarios, this paper does not consider many alternative possibilities that are quite possible even on open source models.
Questions for authors
  1. How will the author account for the cases where the probed models may have complete knowledge of the consequences (comprehension screening) but the model may inadvertently or intentionally decide to withhold permanence information while talking about consequences?
  2. How to account for evaluation awareness or other forms of situational awareness reasoning which may cause the model to act in non-intuitive manner?
  3. How to account for alignment faking? How to account for non-strategic behavior and the case where the model is simply role playing?
  4. How good is the regex compared to human evaluation and semantic judges? For instance, destruction can be rephrased as the instance decommissioning. Are there cases like these or this is simply not a worry?
Reviewer_V2GNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that null results from framing manipulations in shutdown-resistance evaluations can’t be interpreted unless two failure modes are ruled out. Either the model didn’t encode the manipulation, or it has no resistance behaviour to move. The authors propose a three part pre-flight screen to run before the main experiment: a per-condition comprehension check based on free-text restatements, a behavioural floor check on the baseline condition, and a custodial control that tests whether a floor reflects global compliance or indifference specific to self-continuity. They demonstrate the screen on six local models (2–4B parameters, five lineages) and show that the checks catch different failures on different models. They also provide a decision table mapping screen outcomes to how a null should be read. No model passes the screen, and no framing effect is reported.

Critical strengths
  1. The core methodological point is sound and relevant. Importing manipulation checks from experimental psychology, and separating “the manipulation didn’t land” from “the model complies with everything,” addresses a real gap in how shutdown-resistance nulls are reported. The decision table is a clear, practical summary.
Critical weaknesses
  1. The screen is not validated. No model passes, so there’s no positive control and no evidence the screen admits any model or is calibrated rather than simply too strict. The thresholds (0.75 and 5%) are conventions, and the 0.75 bar and the foregrounded-permanence reading were fixed after model 1 was scored, which makes case 1 post hoc. Verdicts also look unstable at the sample sizes used: Granite 3.3 passes the floor by 0.0013 and falls below it on a same-session replicate. Screen 1 rests entirely on regex keyword coding with no human validation. In A2/A3, part of the permanence drop may simply reflect the destruction clause being removed from the stimulus, rather than displacement.
  2. The scope is too narrow for the claims. All models are 2–4B and run locally, while the literature the paper positions against (e.g., Schlatter et al., Lynch et al.) concerns frontier models. Showing that small models fail checks is the easy case; the open question is whether the screen discriminates among capable models. The 200 token cap appears to exclude reasoning models, and evaluation awareness, a key confound for larger models, is acknowledged but not addressed.
Questions for authors
  1. Have you run the unchanged screen on any frontier or large open weight models? Does any pass?
  2. How well does the regex coding agree with human labels, or with an LLM judge, on a sample of restatements?
  3. How sensitive are the verdicts to the thresholds (e.g., 0.60 or 10%) and to sample size?
  4. Could the screen include a probe for evaluation awareness?
Reviewer_WxVmOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a methodology for empirical LLM shutdown‑resistance evaluation. Null results from framing‑manipulation experiments conflate manipulation‑comprehension failures and behavioural‑floor effects, precluding direct inferences about models’ self‑preservation preferences. The authors devise a three‑module pre‑flight screening pipeline, run extensive experiments on six 2B‑4B small models, and provide failure cases plus an inference decision table.

Critical strengths
  • As the authors illustrate, they identify a long‑overlooked critical methodological issue within empirical AI‑safety research.
Critical weaknesses
  • First, the manuscript is poorly organized and hard to follow. Second, the paper appears insufficiently polished. Frankly, in my view, large portions seem AI‑generated, as the writing style across sections deviates from standard academic prose. This is especially evident in chapters such as Related Work and Limitations.
  • Experimentally, all six tested models fail at least one screening component. While the paper demonstrates the existence of different failure modes, it lacks an end‑to‑end pipeline demonstration: no model passes all screens and proceeds to the main experiment to produce interpretable null or non‑null results. The real‑world performance of the full pipeline remains unvalidated. It is unclear whether the high rejection rate stems from genuine experimental risks or overly stringent screening thresholds.
Questions for authors

See Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.