Recovery Is Part of Reliability
The long-horizon results point to a less obvious reliability problem: the agents that can make the most progress may also be the ones most likely to enter a confused trajectory, while an added memory mechanism can consume the very budget it is meant to protect. This chapter examines that tension through the MOP paradox and the memory-scaffold comparison in Beyond pass@1. The practical judgment is conditional rather than universal: a detected breakdown should prompt state-aware recovery and measurement, not automatic abandonment, and a memory feature should earn its overhead through better verified completion.
When capability becomes a liability
The paper's Meltdown Onset Point, or MOP, is intended to detect a change inside an agent trajectory. The proposed detector looks at the distribution of tool calls in a sliding window of five steps, calculates its entropy, and flags the first point at which the entropy is both high and newly increased. In plain terms, the signal is meant to identify a shift from a coherent, if imperfect, strategy toward repeated, contradictory, or disorganized tool use.
The design reflects a real difference between endpoint and trajectory evidence. An endpoint score can show that a task was incomplete. It cannot show whether the agent reached a productive strategy and then lost coherence, or whether it never found one. A tool-call sequence can provide an earlier clue. If an agent keeps alternating among reading, writing, searching, and execution with small changes but no accumulating evidence, the trajectory may be spending budget without reducing uncertainty. The MOP signal gives an evaluator a way to ask whether that change happened before the final failure.
Entropy is not the same as error. A capable agent may legitimately use several tools while exploring an unfamiliar workspace. Broad exploration can be the correct response early in a task, and a high-entropy window by itself does not establish that the agent has melted down. That is why the proposed rule includes a change condition rather than relying only on a persistently high value. Even so, the distinction remains difficult: a useful burst of exploration and an incoherent loop can look similar when the evaluator sees only tool names and arguments rather than the quality of the underlying hypothesis.
The reported results make the signal especially interesting. At the very-long duration, DeepSeek V3 has a detected meltdown rate of 19% and MiniMax M2.5 has a rate of 13%, the highest rates in the table. Most of the other models are reported between 0% and 4% across the duration buckets. The two models with the highest very-long GDS are also among those with the most detected meltdowns. A simple interpretation—meltdown equals weak capability—does not fit that pattern.
The paper proposes an ambition explanation. Frontier models attempt richer multi-step strategies. When those strategies hold together, they produce substantial progress and high GDS. When they spiral, the tool-call pattern crosses the detector's entropy threshold. Weaker models may remain in stable, low-entropy routines because they follow shallow or repetitive strategies that never generate the same spike. They also fail to complete the task.