- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies BERT models using Glauber dynamics, which re-sample one token at a time, to explore what distribution would be sampled by these dynamics over long time horizons. A number of results are derived from this dynamical systems perspective. First, it is shown the transitions are not consistent, so that a stationary distribution is not well defined. Second, results on mixing times are derived, showing that slow and fast mixing regimes exist. Finally, an empirical study was done on the meta-stability of the dynamics, showing that the Glauber dynamics slowly drift through different meta-stable basins reflecting different topics.
Critical strengths
I really like the dynamical systems perspective, and found it a fresh way to study transformer models. Analyzing generative dynamics based on some meta-stability property seems like a cool way to probe the dynamics that goes beyond just looking at the transitions. Theoretical analysis of mixing times is also interesting and novel, as other text generative models don't have a way to naturally consider this.
Critical weaknesses
The sampling results aren't well-motivated because the models studied (BERT) are not used for sampling. And the models used for sampling (e.g. masked diffusion models) were not considered, even though I feel they could be (see below).
Because the motivation for studying these dynamics with BERT is not very tight, the results also felt disconnected and therefore had no joint impact. Each individual result was too under-developed to have a strong payoff. I'll discuss some thoughts about the individual results and then speculate on some ways to strengthen the paper.
Order
The rectangle test / part about the order was not very compelling to me because there have been a rather large number of papers that have explored this idea about inconsistency based on order dependence. While they were focused more on the incompatibility of marginals under different orderings for masked diffusion models, I didn't really see a strong reason you couldn't / shouldn't compare with these. Your case is a special case of MDM where you are choosing order of last token conditioned on others. I was thinking of a recent paper, "DUEL: Exact Likelihood for Masked Diffusion via Deterministic Unmasking", but there are many older papers that discuss that too, e.g. I found "Inconsistencies in MLM" arxiv: 2301.00068. Also: Takateru Yamakoshi, Thomas L. Griffiths, and Robert D. Hawkins. Probing BERT’s priors with serial reproduction chains. In Findings of the Association for Computational Linguistics: ACL 2022, 2022.
There was a comment like "stationary distribution is emergent" - this didn't sound like the right way to express this.
Mixing times
My understanding of this result is that you give an upper bound on mixing time (it would be interesting to have a sense of how tight this is - if n log n is generally required it's a quite important reason this approach should not be useful), and then the example of an empirical situation with the "!" process where you get certifiably slow mixing. I guess I didn't understand the "!" process very well, it seems kind of artificial so I didn't get a sense of whether we should expect slow mixing issues generally or only in special circumstances.
Minor comments: 363 had an unresolved ref.
Meta-stability
The meta-stability / topic result with politics seemed the weakest - both in its execution but also with writing and presentation. One issue is that this became kind of a computational social science direction ("sociotechnical" as written in the paper) - which doesn't really match the rest of the paper. And again, it's hard to see the point of this sociotechnical analysis for a sampling method that nobody uses and the paper is not even advocating as a constructive sampling approach.
In this section Table 3 repeats columns, and it was very confusing to understand what you were trying to show here. "Everything drifts towards politics?" I don't think the results obviously support it, and there's not enough development of the experiment for me to trust the way you measure this.
Minor comments: 462 "5. predicts."
Constructive feedback
I don't know what the remedy is, but I can imagine a few possible paths for how this paper could work, through strengthening the motivation, results or ideally both together.
(a) Rather than glossing over the fact that you are studying a kind of sampling no one uses (sampling via Glauber dynamics on BERT models), maybe it would be better to lean into it. To me, what was fascinating about this paper is the idea that BERT dynamics can be used to sample, in principle, via Glauber dynamics. That means that masked diffusion and autoregressive are not the only two possibilities for text sampling. While it seems from your refs that a few others have noticed this, your analysis can maybe provide some insight into why this doesn't work, and evaluate the promise for whether it ever could. (The n log n result, if it were a lower bound, seems to establish a strong reason that it won't be better than AR which takes only n steps to sample.) So the story would be more like - we have a completely different paradigm for sampling with transformers via Glauber / editing dynamics, could it ever be a dark horse that becomes competitive? Then structure and refine the results to answer this question: mixing time is too slow, no well-defined stationary distribution, leading to biased meta-stable states. And you could speculate (even an experiment here?) about if / how a model could be improved to effectively sample with Glauber. (And instead of barely mentioning GGM paper, you could strongly frame Glauber as a serious Gen AI suggestion and directly critique by using your methods to show why this approach is not promising. )
(b) Another angle that would be interesting to engage with is the recent rise of editing-based masked diffusion models. Llada 2.1 (I think) has an editing model. To me, it is a no-brainer to adapt it to your framework and study if it is qualitatively the same or different to BERT models. I really don't see why any MDM model could not be considered (though the editing models are a more interesting fit for your methodology). There's also been some interesting differences between MDM models trained from scratch and MDM models adapted from AR (like DREAM). Again, I feel your proposed methods could give a unique insight.
(c) If the idea is to only study BERT, and not engage with any recent generative model research, it may be possible but then I think you would need to show that some of these results have bearing on real-world use cases for BERT. You say: "we investigate global semantic behavior through topic-level metastability and transition structure in the induced Markov chain" - but it's not clear how global semantic behavior in practical use cases is connected to your results.
(d) A wildly different direction could build on the mixing time result to ask fundamental questions about when efficient sampling could be done using Glauber dynamics, with some inspiration from https://www.pnas.org/doi/10.1073/pnas.2311810121.
Questions for authors
While I generally feel that the changes in motivation and results needed are too significant to change in the rebuttal, I like the idea for the paper, remain open-minded, will engage with the rebuttals and would change my score with compelling enough arguments.