- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper seeks to understand how language models process inputs that are ostensibly informationally equivalent but are presented in altered modalities (in iconic, tabular, or discursive textual formats).
They evaluate a mix of open and proprietary models on several hand-designed tasks, which require models to answer simple multiple choice questions about stimuli that are presented in the different formats mentioned above. They find that models do not process all modalities with equal degree of accuracy, and that linear probes, by their nature, are unable to fully distinguish different formats.
The ideas in this paper are extremely interesting, but I believe there is a deep lack of conceptual clarity (elaborated upon later) and that there are severe weaknesses in the reporting of results, partly due to those sections being seemingly AI-generated. I believe the work is in need of severe revisions before it is fit for publication
Critical strengths
-
The paper covers a fair deal of conceptual ground with respect to its experiments and I found the problem of using linear probes to distinguish representational formats under simple linear transformations interesting.
-
As language models and multimodal VLMs continue to process ever more modalities of data it is important to understand the extent to which models are sensitive to these difference (if at all). This paper gives preliminary evidence of subtle differences depending on how text input is formatted, even if the informational content is mostly identical
Critical weaknesses
-
I do not believe the different representations are trivially informationally equivalent. They may be equivalent in the eyes of the experiment designers who used the same source dataset to construct the stimuli, but to the eyes of the observing system there may be small discrepancies between formats which, while not necessarily task relevant on the surface, might influence task performance. Since a central claim is that surface form matters (motivation for the tabular condition), if one format makes some kinds of information more easily discernible (e.g., the iconic format for the network diagrams), would we still consider them informationally equivalent? Or does this latter example tap more into the notion of computational equivalence that the authors note?
-
It is argued that the tabular condition is introduced to control for the fact that the diagram input while containing the same information has different surface properties, namely being arranged spatially. I understand the argument the authors lay out in the paper, but am not really convinced that the tables are “a discursive representation with superficial iconic form.” The kind of form they take is highly distinct from diagrams where information is encoded by spatial features and is less flexible in its semantic content than discursive representations due to the tighter syntax. All of which is to say is that I’m not sure that tabular falls between iconic and discursive if the latter two are imagined to be at the poles of a spectrum. Indeed, perhaps some of the lack of differences between tabular and discursive results can be explained in terms of this lack of conceptual clarity.
-
While the opening introduction is largely clear. the bulk of the results section is hard to parse and appears to be 100% AI-generated (as indicated by Pangram <5% false positive rate). The issue isn’t that it is AI generated per se, but that the current wording severely hurts the readability of the paper. I strongly recommend that the authors rewrite this section
-
The presentation of the results is of mixed quality. For example Fig 2 is largely redundant with the tables presented earlier in the paper. The only added information for the by-task family graph is the individual points corresponding to each model, but even that is better represented as grouped bar graphs with error bars (when possible).
-
For the residual stream analyses there are inadequate justifications for why certain probing decisions were made, even while noting the choice of readout makes a difference. For example, “mean-pooling over prompt tokens and reading the last token give different results.” Why would one consider doing one over the other? I can surmise, but this should be explicitly mentioned in the text.
-
There are many results that aren’t discussed or commented on strongly enough given the framing of the paper. For example, what are we to make of the fact that models perform worse on ‘iconic’ tasks where we would expect people to perform better with iconic vs. discursive descriptions. The lack of human baselines also makes it difficult to assess how many of these effects are directionally sound vs. potentially just being noise (the lack of error bars on reported statistics adds to this issue).
-
The pooled model results are helpful to include, but I strongly recommend including the by-model breakdown in the main text (since there is quite a range of models used in terms of size/performance). Additionally for open models, could the authors also include average accuracy with error bars (stochasiticity will be introduced if non-greedy decoding is used)?
-
There are interesting points raised about representational analysis and comparison of different modalities of the same data. Several interesting papers from the field are referenced, which suggest relatively easy to implement strategies relative to linear probes, but these are not conducted.
-
There is an overall dearth of error analyses on, say, the chain of thoughts or even in the patterns of errors that might give a reader insight into why we might be seeing the deficits that we do.
Questions for authors
-
Why do the authors think we see advantages for non-iconic formats for tasks that seem to require visual processing?
-
Could the authors conduct a simple human experiment baseline to ground these results?
-
Could the section on residual stream probes be reorganized into a general section on representational probing and strengthened with non-probing analyses?
Thanks for the comments. Regarding informational equivalence, we believe that the three renderings are informationally equivalent. Informational equivalence concerns whether one rendering can be recovered from another, whereas computational equivalence concerns how easily a reader can draw inferences from each rendering. A rendering that makes some facts easier to discern differs computationally, not informationally, and this is what the study sets out to measure. To make sure that the renderings were informationally equivalent, we ran an additional, unregistered check during the review period: a deterministic parser for each rendering, which rebuilt the generator’s structure exactly from every rendering of every item (§3).
Re the tabular condition: the design requires only that the table state the same facts as the sentences while sharing the surface properties of the diagram. It is not part of our argument that the table lies between the other two formats on a single scale, and we have made this clear in §1.
Presentation: we agree, and §5 has been rewritten, with the scoring terms defined at the outset. Per-model results are now given in Appendix C. The open-weight models used greedy decoding, so single runs were reported; for the hosted models, the scored answer differed on 1.4% and 2.8% of 144 trials run twice. Study 2 has been removed and will be reported separately.
Re the non-iconic formats’ better performance: the iconic deficit rules out reading answers off the arrangement as the models’ dominant strategy, since that strategy would predict an advantage, which was not observed. The small value of reasoning with diagrams suggests, though it does not confirm, the hypothesis that the models fail to transcribe the drawing into a form their reasoning can use; Appendix B reports one such failure. We agree that a human baseline would help ground the comparison, but unfortunately we could not run one during the discussion period.