acceptodds
Withdrawn from ICLR 2027

Identity Is Not the Decision: Certifying Repairs for Frozen Communication Interfaces

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

158 views

Abstract

Frozen communication interfaces can change task meaning without changing message width. Public probes may leave multiple consistent interfaces that permit the same repair, making full identity recovery spend queries on action-irrelevant distinctions. We show how to certify a repair from the shared acceptable-action set across probe-consistent interfaces. We characterize exact nonadaptive query cost by the smallest probe set separating every minimal interface group lacking a common repair, and give a noisy stopping rule with a finite-sample bound on unsafe certification. Across learned-agent pairs, anchor-consistent messages induce a common receiver action on percent of held-out states; uniform selection among them has a percent empirical upper bound on mismatch with the frozen receiver. On calibrated noisy channels, stopping uses fewer probes than target-symbol or full-interface identification and abstains when probes are uninformative. A complete automata catalog shows query savings against standard and advice-enhanced learning. On a message-dependent task, repair restores reference-level return, and released-protocol audits evaluate held-out workflows.

Primary area:
reinforcement learning
Keywords:
multi-agent reinforcement learning, emergent communication, black-box auditing, partial identification, decision certification, active learning, automata learning, model repair

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7SpcOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper aims to address the issue of communication mismatch between two agents. The authors argue that, rather than fully reconstructing the receiver’s entire “lexicon,” it is sufficient to identify and verify an “action” that remains mutually acceptable to both parties under all possible interpretations. The proposed approach was tested through simulations in a controlled, local environment.

Critical strengths

S1. Given that the task and operational space are already defined, it is more reasonable to validate a proposed fix by ensuring it yields equivalent decision outcomes, rather than attempting to comprehensively specify the interfaces.

Critical weaknesses

W1. According to my understanding, the proposed method requires knowing all possible dictionaries and error types in advance. and it is impractical in real case with massive cost.

W2. Lack of strong motivation and empirical benefits. When tested in a second environment, the proposed fix completely failed to improve task success.

W3. Writing should be improved substantially. The content in the main body is exceptionally difficult to follow.

Questions for authors

Q1. In the experiments, exactly how many total queries were used when you count the massive upfront testing (∣B∣⋅d\vert{}B\vert{}\cdot d reads) needed to set up the proposed method?

Q2. Why was the safety validation path disabled in local MCP tests?

Reviewer_3AmSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how to repair frozen communication interfaces without fully identifying them. The proposed approach selects a repair action that is acceptable under every interface consistent with the observed probes. The paper characterizes the required number of probes and introduces a stopping rule for noisy observations. It evaluates the approach on 225 learned sender–receiver pairs, alongside synthetic experiments and protocol-based evaluations.

Critical strengths
  1. The distinction between identifying an interface and identifying a sufficient repair is useful and clearly motivated.
Critical weaknesses
  1. The practical benefit is limited in the main learned-agent experiment. Section 5.5 reports that agreement with the reference receiver increases from 58.5% to 64.8%, while task success remains approximately 16.7%. Thus, better reproduction of the receiver’s behavior does not translate into better task performance in this setting. The separate message-dependent experiment demonstrates improved return, but the paper should explain more clearly why the main learned-agent experiment shows little improvement in task success and under what conditions the proposed repair is practically useful.
Questions for authors
  1. Why does improved agreement with the reference receiver produce almost no improvement in task success in the main learned-agent experiment? What differs in the message-dependent experiment that allows repair to improve task performance?
Reviewer_7CNtOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is focused on emergent communication, and defines a setting where two agents are communicating via a known codebook, and a new agent takes the place of one that does not have knowledge of this codebook. The question that seems to be examined is then how communication can be effectively reestablished and under what conditions.

Critical strengths

There are no points in this paper that make me believe it should be accepted in current form.

Critical weaknesses

Overall, much of the writing appears Ai generated and I found it extremely hard to follow the paper or grasp it's impact. There is a small chance I am missing something as I am not familiar with this area.

Examples of writing that is extremely hard to follow; In the first sentence - "Frozen communication interfaces can change task meaning without changing message width. “task meaning” and “frozen communication interface” are not defined.

In the second sentence "These checks" - nothing has been introduced as a check.

"A second task environment. Across six checkpoints and 36 pairs…” - the environment is never named.

This continues and is pervasive throughout the manuscript.

Specific examples of AI-sounding text: "The protocol-server audit reads published package versions installed locally and queries each server over its own standard input and output channel. Each probe is a read-only metadata request." (top of reproducibility statement).

“The object certified here is the repair action the evidence leaves, and its price is the size of the set that evidence fixes.” - related work, page 3

"What a rule reads bounds what it can certify, and reading anchors that separate the two candidates is what removes the case. This is why the axis the evaluation sweeps is the budget.” - after prop 2. The second sentence doesn’t follow.

It is prevalent in many of the headings which are conversational in nature and is especially pronounced in the appendix;

  • "B.1 THE GATE TRADES COVERAGE FOR SUPPORT"
  • "B.3 AN ENDPOINT THAT SHAPES ITS OWN PUBLIC SURFACE"
  • "B.6 THE INSTRUMENTS, STATED"
  • "B.5 WHAT THE MESSAGE CARRIES"
  • "The radius is the whole story." - bottom of page 24.
  • "The full filesystem chain, and where the calibration runs out." - bottom of page 27
  • "The repair literature, work by work." - top of page 44

There are many other common AI writing patterns; for example there are almost no introductory sentences that ground the section within the paper and serve to orient the reader - instead simply jumping in immediately from the heading. For example;

"M.25 WHICH END PERFORMS THE REPAIR. A repair can be installed at either end of the channel: the sender can transmit the message the receiver needs, or the receiver can overwrite the message it received. The two are the same intervention whenever the interface is invertible, because in that case the receiver’s effective input is the installed message either way. That premise is measurable, and it was measured:"

"I IMPLICATIONS; Cost follows the target. The decision-level view also predicts the shape of the cost curve, and the prediction holds where the target restricts the goal set"

Very specific non-widely known properties are introduced without citation or explanation (for example 2-helly, time-uniform confidence sets)

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.