- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper asks whether you can control a frozen model, with no fine-tuning and no extra input slot, just by changing statistical properties of its context, such as the sign of a correlation or the variance of the data. Using PFNs, trained on data the researchers fully control, the authors show that such carriers can switch which computation the model performs, with graded control among up to eight states. A carrier's meaning is set entirely by the training distribution: equally informative carriers can take up to 12× longer to learn, the model reads carriers as evidence under its training assumptions, and it suppresses state combinations it never saw. In pretrained LLMs, statistical properties of few-shot examples reliably shift behavior, but these properties mostly can't be given new meanings within a prompt, and whether they can depends heavily on exactly how they appear.
Critical strengths
-
I like that the paper separates how much information a statistical property carries from how easy it is for the model to learn. The authors pick 48 carriers in advance and calibrate them so that an ideal decoder recovers the state equally well from each. Even so, learning time still varies about 12x, and a few structural features fixed before training explain most of that. Without this controlled setup, the point would only be an intuition.
-
The experiments are carefully designed and the claims are well scoped. Control is tested by changing only the carrier while keeping everything else the same, and results are compared against exact Bayesian references. The authors are also clear about what they don’t show, e.g., no claims about internal mechanisms and no execute-level results in LLMs, so I found the main claims easy to trust.
Critical weaknesses
-
PFNs are trained to approximate the posterior under their training distribution, so it's not surprising that carrier meaning comes from training, that predictions track the training posterior, or that unseen combinations get suppressed. The authors acknowledge this, but Section 3 is still presented as a main finding. The off-training-strength test (Bayes_train vs. Bayes_true) is the most informative part, and I'd like the paper to say more clearly which results go beyond what the training objective already predicts.
-
It is not clear what the "interface" framing adds beyond existing work on in-context task inference. If we remove the new terms (carrier, writable, READ/EXECUTE, bindability), much of the paper looks like a model inferring a latent task from context, which has been studied a lot. The paper would be stronger if it showed at least one result that the interface view predicts and those accounts do not.
-
The strong claims (execute-level control, Bayesian semantics) are only shown in small PFNs. In LLMs, the paper shows sensitivity to statistical properties, which is already well documented, plus limited and setup-dependent binding. A lexical cue binds in 12/12 models while most statistical properties do not, so it is also unclear when someone would choose a statistical carrier over a plain instruction. I would suggest the authors either narrow the claim in the abstract or add execute-style evidence in at least one pretrained model.
Questions for authors
-
Which of the PFN results in Section 3 do you think go beyond what the PFN training objective already predicts? Could you point to one result that would have come out differently if the model were just approximating the training posterior?
-
Is there any result in the paper that the interface view predicts but existing accounts of in-context task inference would not?
-
For the LLM experiments, is there a setting where a statistical carrier works better than a plain instruction or lexical cue? If not, what is the practical use of statistical interfaces in pretrained models?
See:
Xie, S. M., Raghunathan, A., Liang, P., & Ma, T. (2022). An explanation of in-context learning as implicit Bayesian inference. ICLR.
Chan, S. C. Y., Santoro, A., Lampinen, A. K., Wang, J. X., Singh, A. K., Richemond, P. H., McClelland, J. L., & Hill, F. (2022). Data distributional properties drive emergent in-context learning in transformers. NeurIPS.