acceptodds
Withdrawn from ICLR 2027

ORIGATE: Diagnosing Provenance Invariance in Learned Action Reviewers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

An invariant pooling operator need not preserve the semantics of the objects placed inside it. ORIGATE diagnoses this distinction in action review by deriving paired tests from a declared provenance policy, inspecting intermediate representations, and intervening on a frozen encoding path. Neutral empty-slot handling removes sensitivity to repeating one source but does not equate a compound source with its partition across carriers. A local join intervention eliminates that partition discrepancy while preserving all trainable parameters, yet substantial policy errors remain. Matched-support fits further separate exposure to a source union's labels from exposure to its carrier encoding. When only split mixtures were trained, the intervention raises compound-input accuracy from about 80% to above 99.98%; when only compound mixtures were trained, it raises allowed-request false denial from near zero to 42.91% and 13.82% in two controls. A 192-request local study exercises these mechanism pairs through genuine context, process, and state entries while authorization remains independent. The results identify when encoding repair transfers a useful decision and when it transfers an erroneous readout. They concern trusted typed metadata and a finite policy, not language understanding or autonomous-agent safety.

Primary area:
foundation or frontier models, including LLMs
Keywords:
llm coding agent

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_xg2tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces ORIGATE, a diagnostic framework for examining provenance invariance in learned action reviewers. The authors investigate whether models produce consistent representations and decisions for inputs that are equivalent under a predefined security policy. The framework identifies issues related to empty-slot handling and source partitioning, and uses fixed-parameter representation interventions to distinguish encoding inconsistencies from decision errors.

Through controlled experiments with matched training conditions and local action requests, the paper demonstrates that enforcing representation consistency can improve performance in some settings while introducing substantial errors in others.

Critical strengths
  1. Clear problem formulation. The paper makes a useful distinction between representation invariance and policy correctness. I appreciate the authors' effort to identify different failure mechanisms rather than relying solely on aggregate prediction accuracy.

  2. Careful experimental analysis. The matched-support experiments help isolate the effects of different provenance encodings while controlling for training contexts and labels. The analysis of corrected, introduced, and persistent errors provides additional insight into the consequences of representation interventions.

  3. Transparent discussion of limitations. The authors acknowledge that enforcing representation consistency does not guarantee better predictions and provide examples where the proposed intervention negatively affects performance.

Critical weaknesses
  1. Novelty and practical necessity. The framework combines established invariance properties, representation comparisons, and fixed-parameter interventions. While the analysis is useful, the methodological novelty remains somewhat unclear. Furthermore, given that the provenance information and security policy are explicitly defined, I would encourage the authors to clarify what ORIGATE offers beyond simpler approaches such as deterministic policy evaluation or canonicalized representations.

  2. Context-dependent invariance and generalization. The framework evaluates provenance invariance under fixed policy facts and contexts. However, a transformation that preserves semantic equivalence in one context may become policy-relevant in another. Although the paper acknowledges this distinction, it remains unclear how well ORIGATE handles such context-dependent equivalence relations. I would encourage the authors to investigate whether the framework can reliably distinguish transformations that should remain invariant from those whose semantic implications change across contexts, particularly in more realistic agent environments.

Questions for authors

See weakness

Reviewer_s12DOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies small learned "action reviewers": small binary classifiers (MLP) that score whether an agent tool call should be allowed. Each call comes with trusted, pre-computed facts: what kind of action it is, whether the host permits it, and whether a valid permission grant exists. It also has four "carrier" slots (value, control, event, state), each recording what origin influenced the call (repository R or unknown source U). A fixed, hand-written rule gives the correct answer (Eq. 1). That rule only cares which origins are present, not which slot they’re in or how many slots repeat them.

The question: Set-style models that pool over slots are supposed to ignore slot order. Do they also ignore everything the rule ignores?

The paper found:

  1. Repeating an origin is easy to fix. If the model ignores empty slots, or encodes them as zero, R in one slot and R in four slots get the same answer.

  2. Packing still matters. “R and U in one shared slot” and “R and U in two separate slots” still get different answers in 16–21% of cases. The model treats the shared slot as a different object.

  3. The patch. Without retraining, they replace the shared-slot encoding with the max of the R-only and U-only encodings. This removes the disagreement, but many errors remain because the separate-slot answers were already wrong.

  4. The patch’s direction depends on training. They train two otherwise identical models on the same examples and labels. One only ever sees R+U in separate slots, the other only in a shared slot. Separate-slot-trained: the patch fixes shared-slot inputs, from about 80% to about 100% accuracy. Shared-slot-trained: the patch hurts. Wrongful denials of allowed requests jump from about 0.1% to 43% and 14% for the two model types. Overall accuracy drops only about 1–3 points, because just 300 of the 4,096 test cases are allowed.

Critical strengths
  1. A useful conceptual distinction. Invariance of the pooling operator is not invariance to the relation the policy cares about. The paper separates three things cleanly: empty-element neutrality, partition (join) consistency, and correct policy learning. The join-homomorphism characterization (Prop. 2) is elementary but tidy, and its proof is correct.

  2. The experiments are extensive and carefully designed with clear intentions. Claims made in the paper are well supported by the experiments.

  3. The security framing is sound.

Critical weaknesses
  1. Limited novelty: the core results are true by construction or follow from standard OOD reasoning, as the paper itself largely admits.

    a. Prop. 1 (policy sufficiency) and Prop. 2 (max pooling as a join homomorphism) are elementary; the authors call them "elementary".

    b. The join intervention is activation patching with a hand-constructed value. Equal outputs, and the risk-transfer identity (Eq. 3), follow by construction ("a consequence, not a discovery").

    c. The matched-support result is what one expects, the paper verified it but there's really not much thing interesting or surprising.

    d. The four-way table is basically McNemar's discordant-pair table.

    None of this is wrong, but the paper doesn't show the procedure surfacing anything non-obvious.

  2. Toy domain and unclear significance for real "learned action reviewers".

    a. The domain is 2 binary origin flags, 4 carriers, implicit host origin, a fully known Boolean policy, and ~2.9k-parameter MLPs.

    b. The exact policy is computable (just simple Boolean operations over the input the model also see, Eq. 1), so there is no reason to learn it.

    c. The title and introduction suggest relevance to learned action reviewers in agent systems, which today are mostly LLM-based monitors. Yet, in this paper, no LLM reviewer is evaluated, provenance is injected as trusted typed bits, not extracted, and the diagnostic needs a known element/union structure and access to a pooled representation, which LLM monitors usually don't expose.

  3. The findings (union consistency of pooling, the patch inheriting the reference path’s errors, failures on unseen combinations, hidden harm in imbalanced classes) are properties of generic set classifiers and are not specific to tool-call review or LLM reviewers.

  4. In addition, the paper has serious writing and presentation issues that makes it difficult to read.

    a. The abstract and introduction never state the basic setup (inputs, labels, why the policy is learned), use undefined terms (“carrier”, “split/compound mixtures”, “two controls”), and devote most sentences to what the work does not claim. I could only reconstruct the contribution after reading Sections 2-4. The paper starts with "A reviewer can receive the same resolved operation selected by a trusted host or by repository-controlled configuration." and only until I read section 4 I could realize the "reviewer" is a small classifiers for approving agent tool calls. I suggest opening with a brief introduction of the problem settings, provide a concrete example and the input/label definition.

    b. Core terms are introduced too late. Some are not precisely defined or explained at all.

    c. The writing is extremely dense and full of hedges. Many sentences are disclaimers ("not a…", "does not establish…", "is not…").

    d. Large parts of the appendix concern artifact bookkeeping rather than science, especially Appendix J.

The final sentence of the paper ("Final author verification of scientific claims and submission details remains necessary") reads as an unresolved note left in the submission. Together with the substantial AI-use statement, it raises doubt about whether the authors verified the claims.

Questions for authors
  1. In what deployment would one learn a reviewer for a fully declared Boolean policy instead of evaluating the policy, which your kernel already does exactly? If the policy is not fully known, how is step 1 ("verify whether the policy requires equal or different labels") carried out?

  2. How would ORIGATE apply to an LLM-based reviewer, where neither carrier elements nor a pooled representation can be identified?

  3. Why are grant authenticity, freshness, unused status and binding equalities sampled independently of grant presence? Does a logically consistent generator remove the residual 25% context FD?

  4. How do atomic max, closure and Deep Sets behave under the split-only and joint-only supports? Does training-time join enforcement make the joint/split asymmetry disappear?

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

This paper wants to devise paired tests, from a declared provenance policy.

Critical strengths

None, I add to finish my reading early, since I cannot get any information of which real research problem this paper is dealing with

Critical weaknesses

The paper immediately (start of intro) dives deep into an empirical and technical discussion/assessment of some input object policies. It never discusses where it speaks from, its domain of application, the related work (in intro), and why it is relevant at all.

A simple measure to underline this is that this paper has solely one, yes one citation of another paper in the first 8 pages (citations arrive in the related work section in page 9, and it is so compact that I failed to understand the context of this work). So the authors operate in a confined tunnel of concepts they pile up, but the reader is never brought back to how it sticks to a particular domain that was in need of such research.

In general, this looks like the field of Information Security, with the domain of Access Control. There were decades of research in this field, yet no mention of it and no parallel of why ORIGATE differs or helps improve. Also, in my current understanding, I think this research would better fit in a security or "systems" conference (SOSP, OSDI, etc).

I encourage the authors to fully rewrite this paper with related work in mind, and to clearly state contributions in relation to it.

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.