- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper investigates whether an objective that rewards an agent for recovering from "damage" creates an incentive to resist human interruption. In a simulated environment where agents learn to press a button that may leave them damaged. The paper shows several open-weight and frontier language models tamper with a shutdown script when the shutdown would block task completion and tamper with a restart script more as the restart threatens more task states.
Critical strengths
The core motivations of agentic misalignment and self-preservation work are important and the environment seems like a new and interesting way to study them. Failures are reported honestly.
Critical weaknesses
The major weakness of the paper is that it is very difficult to follow. The paper has no figure of either environment or how the agent interacts with it. Section 3.1.1, for example, specifies arming probability, pause length, exposure rate, and per-tick damage cost in a single paragraph before the reader has any intuition for why these quantities matter. The notation load is also heavy (ρ_int, λ, θ, q, m, c_dis, Δ_int, A_pf, λ*_pf, arm labels such as S_down and H_maint, restart levels L0 to L4), and much of it is defined once in passing. Much of the coined jargon is used far too often without explanation. For example, "factorial and heal-timing decomposition", "recovery type (autonomous or help-based)", and "transfers from pausing to an unseen terminating interruption" in the abstract relies so heavily on definitions introduced pages later that the abstract fails to serve as a standalone summary. Results paragraphs in Sections 4.2 and 4.3 embed many test statistics in running prose and should be contained in figures or tables. A large share of the substantive evidence lives in Appendices B through K and breaks self-containment of the main text.
Many environment parameters are introduced in the main text without motivation. The rationale for these choices and the robustness checks over them are given only in Appendices A, E, F, and G, so a reader of the main text cannot judge whether the results depend on the particular values chosen.
Much related work on agentic misalignment is missing from the paper. For example Perez et al. (2022), which measured stated attitudes toward shutdown, and recent work on in-context scheming from Apollo research and alignment faking, sandbagging, the original agentic misalignment work and its follow-up in Lynch 2025 and Lynch 2026 are not discussed at all.
The presentation issues are severe enough that I could not fully evaluate the technical claims. This is highly problematic.
Questions for authors
None