acceptodds
Withdrawn from ICLR 2027

Functional Portability, Geometric Limits: Cross-Construction Number Interventions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-construction transfer reveals reusable output control, but which properties of the underlying mechanism does it identify? We fit grammatical-number directions on subject–verb agreement and apply them to disjoint reflexive outputs. Across 21 held-out Pythia size–seed models, all three target constructions reach persistent onset (63/63 target trajectories), versus 0/63 for a fixed orthogonal control; this original comparison uses unmatched edit amplitudes. A separate equal-edit-norm follow-up on exposed original materials retains an advantage over random and orthogonal directions, but not a uniform advantage over a simple lexical-number direction. Further boundary tests show that selected supports concentrate the effect while broader random supports recover much of it, and that changing the writing site changes a fixed direction's structural response. On a prospective BLiMP subset, target gain meets its requirement but joint collateral preservation does not; lowering dose reduces both gain and measured collateral. Supplementary reorientation on exposed materials offers no demonstrated advantage over a diagonal approximation. Together, these studies distinguish portable number control from evidence for a uniquely syntactic, compact, position-invariant, or selectively low-collateral mechanism.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, causal interventions, grammatical number, training dynamics, representation transfer

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_UvigOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies what cross-construction transfer of grammatical-number interventions reveals about language model representations. The authors find that number-related directions transfer across grammatical constructions, but further experiments show that these effects do not necessarily establish a unique syntactic mechanism and depend on coordinate selection, intervention position, and effects on other outputs.

Critical strengths

None

Critical weaknesses

Despite my background in related research, I found this paper extremely difficult to understand. The writing relies heavily on unnecessary jargon, complicated terminology, and excessive qualifications, making it difficult to identify the main contributions and follow the arguments.

The lack of clarity prevents me from independently evaluating the experimental design, the relationship between the evidence and the claims, and the significance of the contributions. A fundamental purpose of a scientific paper is to communicate research so that other researchers can understand, critically evaluate, and verify its findings. I believe this paper fails to fulfill that purpose.

Questions for authors

I do not have specific questions for the authors because my primary concern is the overall clarity and readability of the paper, rather than any particular technical detail. I believe the paper would benefit from a substantial rewrite before its scientific contributions can be properly evaluated.

Reviewer_jqRNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper asks whether a direction for grammatical number found in a small language model transfers across syntactic constructions. The authors fit a difference-of-means direction on subject-verb agreement sentences, e.g. is / are, at one layer and one token position. They then apply it as an intervention to sentences whose output is a reflexive pronoun, e.g. himself / themselves, in three constructions with a distractor noun. Evaluated on a set of independently pretrained Pythia models, the intervention flips the number of the reflexive in every model and construction, while an orthogonal control fails to do so.

At the same time, the authors argue that this result does not support stronger readings of the transfer mechanism, since: 1) A direction fitted on bare noun phrases performs as well as one that was fit on subject-verb agreement sentences, 2) The same mechanism fails to fully transfer to other model families such as OLMo, 3) In the Pythia family, the effect is concentrated on about 8% of coordinates, but is matched in effect by a random 33% subset, 4) Writing the same direction to another token changes its effect, 5) On a small subset of BLiMP, the intervention is shown to cause substantial unintended changes to model outputs. The authors thus conclude that interventions can steer model output w.r.t. grammatical number, but with substantial limitations to the interpretability of the transfer effect.

Critical strengths
  1. Thorough experimental design with respect to controls and ablations, that prioritizes analytical depth and root-causing of the identified mechanism, even where the results are inconclusive. For each alternative explanation of the identified transfer effect, the authors include a control designed to rule it out, and the pass / fail rules are shown not to depend on the exact control hyper-parameters, which suggests robustness of the presented findings. Abandoned studies and a bug found during the project are also reported rather than omitted. Even though this work leaves a lot of open questions, it is self-critical with cautious, rather than exaggerated claims, which is fundamentally commendable.
Critical weaknesses
  1. The paper is very hard to read and parse. Each finding is stated inside a paragraph of qualifications, so the reader has to track down the core result before judging it, and the main text cannot be followed without the very long and dense appendix. The authors seemingly use their own terminology and internal jargon without ever defining it upfront. The consequence is that a small set of meaningful findings reads as an undifferentiated, sometimes inscrutable list of caveats. I would suggest rewriting the main text around the main transfer result and the two comparisons the core argument of the paper depends on, i.e. the noun-only direction and the side-effect test, stating each finding once in plain language before qualifying it, and treating everything else as secondary, supplementary findings.

  2. The contribution is narrow and would be better suited for an interpretability workshop. All positive findings come from small Pythia models, the intervention method itself is borrowed from prior work, and the authors themselves constrain their novelty claim to the general lesson that transfer does not imply a mechanism. The one attempt on another model family at the same scale found nothing, and the paper does not investigate why, so the reader cannot tell whether its the transfer effect or the experimental protocol that failed to generalize. The authors could address some of these issues by rerunning the main study with the noun-only and size-matched controls included from the start, adding at least one larger LLM of 7B parameters or more to strengthen its relevance to practical GenAI applications, and diagnosing the failure on the second model family before presenting it as a result.

Questions for authors

The three families outside Pythia give three different outcomes - SmolLM2 shows the full effect and loses the small-subset version, Qwen2.5 reproduces one diagnostic, while OLMo-2 shows nothing. Can the authors offer some diagnosis as to why this is the case? This would strengthen the general applicability of the proposed methodology since, as its stands, it does not appear to extend beyound select model families.

Reviewer_fYKSOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors find that the LLMs exhibit number-agreement between different linguistic constructions (cross=construction transfer). For e.g. the number from a subject verb agreement transfers to reflexive pronouns. They show evidence from causal interpretability experiments to prove this.

Critical strengths
  1. Interesting question around whether grammatical number directions are universal in a model.
  2. Causal interpretability methods seem to have been used.
  3. large scale evaluation design (21 Pythia models) although limited to one family.
Critical weaknesses

Very poor writing, likely AI slop word-salad. Starting from the abstract, there is excessive use of jargon to motivate the problem as well as communicate the results. Really hard to follow most of what's written in this paper. Though the problem itself is elegant and simple, the communication is terrible (100% AI generated abstract according to Pangram). These are the first few lines of the abstract. It's unclear what they even mean to a non-linguist (or even a linguist for that matter), and can be introduced in more simple language . Here's a set of phrases/sentences from just the abstract that make no sense:

  • "Cross-construction transfer"
  • "relevant output-control"???
  • "separate equal-edit norm follow-up"
  • "retains an advantage over random and orthogonal directions but not a uniform advantage over a simple lexical-number direction"
  • "boundary tests"
  • "boundary tests show that selected supports concentrate the effect while broader random supports recover much of it"
  • "On a prospective BLiMP subset target gain meets its requirement but joint collateral preservation does not"

While I can gauge what the results are, I really don't see the point in reviewing a paper that's so completely auto-generated by an agent. Will let the AC decide if the paper should be reviewed in its current state.

Questions for authors

N/A

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.