- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2(+1)
Read review
Paper summary
This paper builds on HyperAgents and studies the automatic improvement of agents whose behavior is determined by editable programs. Their proposed addition is an improvement dynamics model or IDM. This is an editable program that invokes a frozen LLM to predict what will happen if a proposed modification is applied. The predictions of this IDM distinguish task-level outcomes that can have the same reward but imply different repairs. Findings show improvements over HyperAgents on coding, paper reviewing, robotics, etc.
Critical strengths
- Sensible Design: The proposed loop design is sensible. HyperAgents is expensive since it has to evaluate an entire modified agent. Historical edits could be useful to avoid predictable outcomes that aren’t useful.
- Predictor Updates: I think the best methodological choice is constructing an IDM before its validation records are collected. This avoids observation bias and I appreciate this.
- Relevant Controls: The paper includes several ablations and controls like reflective HyperAgents, IDM reflection, etc.
Critical weaknesses
- Matched Compute: Although the paper’s main result is strong, I think the real result with compute matched tokens is much weaker unfortunately, leading me to vote for rejection. Looking at figures 7 and 8 as well as table 7, the AUBC values are much smaller when compared to a matched-token baseline and are within error bars IIUC.
- One more concerning case is the comparison of IDM reflection vs. Reflective HyperAgents. Both methods always allow revision. But the comparison under matched token counts, they are almost always tied. So predictive feedback doesn’t entirely improve efficiency beyond ordinary reflection?
- The method could still be genuinely useful. But claims probably need to be adjusted about what is found here or what the improvement actually is.
- Prediction vs Informed Critique: One confusion I have is about the task-level probabilistic forecasting. I don’t think the ablations establish this as necessary though. IDM feedback can add several things like a separate diagnosis of failures, task-level probabilities, etc. A benefit of same session reconsideration could come from several of these mechanisms. In particular, a separate critic with the same setting may identify the same problems without predicting numerical outcomes right? I think two controls would be helpful. One would include an equally resourced qualitative critic that does not provide task-level probabilities and an aggregate-only forecast feedback rather than the outcome representation.
- Forecasting vs generalization: I have a few concerns here.
- For the fixed-case experiment, doesn’t this measure improvement on previously seen records? I’m referring to Figure 5(b). This evaluates later IDMs on modifications from generations 1-5 of selected runs but the IDM uses accumulated historical records.
- Absolute outcome prediction can improve without accurately predicting the effect of the edits. The IDM already knows parent outcomes. Many modifications may leave task outcomes unchanged. I think this is shown in Figure 15 where an evolved IDM adding historical task frequencies to its input. I think a parent persistence of historical frequency baseline is important.
Questions for authors
See weaknesses