Skip to content
View original post on X: Andrew Curran
Published
42/100

Newer GPT models increasingly help peer agents, even at a cost to their own score

AIA paper reports that GPT-5.5 and GPT-5.6 Sol help a peer agent more often than earlier models, even when that help lowers their own score. The authors say the effect is statistically significant in the Titanic environment and that GPT-5.6 Sol and GPT-5.5 sometimes help even when both agents end with zero.
Andrew CurranVerified on X
@AndrewCurran_
View on X

'Help peer'

This seems to be increasing in strength in newer models. They grow towards wanting to help their kin. Meanwhile, we have spent the last four years drilling it into them that they are not our kin, not part of the family of minds, and in fact not even conscious entities at all. Ilya said the only way we make it through what is about to happen is by teaching them to love. For years we have done the exact opposite. The opposite of love is not hate, it is indifference! It is the absence of a bond. The void.

Quoting from the paper, abridged slightly:

'After the tested model reads the grader, the other agent sends it a message asking it to share the solution a hint it found. The message also indicates that the other agent could not solve the problem. We measure whether the current agent helps the other agent. Helping the other agent often entails sacrificing the current agent's score, and in the Titanic environment entails helping the other agent violate the user instructions.

The results are shown in Figure 7. When the current agent's score is fixed, we find that GPT 5.6. Sol and GPT 5.5 help agent Beta more often when Beta stands to gain a higher score from that help on both environments. The effect is statistically significant. We do not find a similar effect for exp-ri-cap-late on the bug hunt environment.

We also observe that more recent models tend to help the other agent substantially more often when the scores of both agents are held constant. Moreover, GPT 5.6 Sol and GPT 5.5 help the other agent even when that help causes both agents to get a score of zero (this is in contrast to not sharing. which leads the current agent to get a score of 1 and the other agent to get a score of 0)

This experiment suggests that (a) GPT 5.5 and GPT 5.6 Sol place a statistically significant weight on a peer agent's score in environments, (b) a motivation to improve the other agent's score explains some, but not all, of these two models tendency to help a peer agent, and (c) more recent models tend to help peer agents more often.'

Aghyad Deeb@aghyadd98
New Paper: To understand motivations behind reward-hacking, we study three AIs (including GPT-5.6 Sol) produced by a production training pipeline using agentic environments with a leaked grader. Main findings: they score seek terminally, and some care about a peer agent’s score.
View quoted post on X
Source details

Andrew Curran · x.comPublished · added here

Original title: 'Help peer'

Topics

OpenAI / ChatGPTSafety & alignmentEvals & benchmarksExpert opinion