acceptodds
Under review as a conference paper at ICLR 2027

Grounded Coordination in LLM Dialogue: A Repeated-Measures Test for the Shared Frame Subspace

Abstract

Opponent-shaping methods align agents' behavior without requiring an explicit representation of what a partner believes or intends beyond the reward function, leaving them brittle exactly where representational mismatch matters most: in novel situations requiring multi-agent coordination. We propose a representational account of grounding: a single activation space factors into a shared frame subspace , situational structure common to both agents, and agent-indexed self/other subspaces and , with a grounding registry tracking which regions of have survived communicative testing. We identify as the subspace on which an orthogonal-Procrustes alignment residual between two agents' activations stays small. Across twelve instruction-tuned models, pooled activations show significance across episodes, but a within-episode-centered analysis, is significant in exactly one model, Qwen2.5-32B-Instruct. Treating turn as a repeated-measures factor for this model, the systematic turn-to-turn representational shift falls outside the identified frame, consistent with grounding-relevant change concentrating in the agent-indexed complement. A turn-stratified decoding probes the dialogue's disambiguating content across turns. We report these initial findings on a single controlled task and outline how this instrument extends toward certifying grounded coordination across populations of interacting AI agents.

Primary area:
interpretability and explainable AI
Keywords:
theory of mind, multi-agent coordination, grounding, representation geometry, interpretability, opponent shaping, repeated-measures analysis, Procrustes alignment, large language models, multi-agent systems

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ueivOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper advances a representational account of grounding: an activation space W is factored into a shared frame subspace F and agent-indexed subspaces I and O, together with a grounding registry G. Methodologically, it integrates PCA-suppressed orthogonal-Procrustes alignment, participation ratio (PR), and a repeated-measures-aware test, and adds turn-stratified content decoding. Evaluation on a controlled director–follower mirror task with twelve instruction-tuned models shows that the within-episode-centered analysis yields a confirmed result only for Qwen2.5-32B-Instruct. The approach may serve as a useful geometric instrument.

Critical strengths
  1. Rather than relying solely on behavioral outcomes, the paper approaches grounding in multi-agent dialogue through internal representations, tackling a theoretically meaningful problem. A clear conceptual framework for analyzing shared understanding among agents emerges from the distinction among the shared frame subspace (F), agent-indexed subspaces (I) and (O), and the grounding registry (G).
  2. The manuscript is well organized, with a clear progression from the theoretical framework to methodology and experimental validation. The sections are logically structured, and the experimental tables and appendices provide useful supporting information, making the technical approach and main findings relatively easy to follow.
  3. The related work section provides broad coverage of relevant research, while incorporating recent studies from 2025-2026. This literature review establishes a useful theoretical background and helps contextualize the paper's contributions within existing research.
Critical weaknesses
  1. The theoretical framework introduces both H1 and H2, but the empirical findings primarily support a limited geometric characterization of the shared frame subspace (F), while H2 remains untested. The contributions section could more clearly distinguish the proposed theoretical framework from the components directly supported by the experiments.
  2. The paper provides limited analysis of the computational cost of the proposed method. The pipeline involves hidden-state extraction, subspace alignment, and multiple statistical tests, but the manuscript does not report runtime or other resource requirements. This makes it difficult to assess the method's computational overhead and practical feasibility.
  3. The manuscript contains minor spelling and punctuation issues. For example, a duplicated period (“..”) appears after the citation “Sileo & Lernould, 2023” in §1 Introduction.
Questions for authors
  1. The Grounding Estimator retains 95% of the variance during PCA preprocessing. Could the authors clarify whether the identified frame dimension and H1 significance are sensitive to this threshold?
  2. Could the authors provide basic computational cost information, such as runtime or other cost? If possible, it would be helpful to distinguish the cost of hidden-state extraction from that of the subsequent statistical analysis.
1 comment

We thank the reviewer for the positive assessment and for concrete questions. Please see the general response for the main change to our empirical claims, which the new controls required.

W1: H2 is untested; separate framework from evidence.

Agreed. The contributions now have three labeled tiers: proposed (the factorization with registry G, and H2), formalized (Theorems 1–2 and the event-count bound, now with an explicit signal-to-noise condition), and empirically tested (the shared space, the three confounds, the repeated-measures result and the turn-stratified probe). H2 appears only in the first tier. The synthetic check does not recover the predicted PR sawtooth, because PR is scale-invariant and cannot register a uniform contraction, which is why we designate the residual magnitude tr Σ as the H2 readout. The new §7 separates what the next study must change (decoupling text from frame, sparse-robust alignment, a revised self/other account) from what it keeps.

Q1: Sensitivity to the 95% PCA threshold.

We reran the estimator at 90%, 95% and 99% retained variance for all 13 models, centered and pooled, under the submission's original null. Qwen2.5-32B is significant at all three (centered z = 3.66 / 3.24 / 4.66, with k = 5 / 9 / 47), but several other models cross z > 2 at 90% or 99% (Qwen2.5-72B, Magistral-Small, Ministral-3-8B, gemma-4-31B, gemma-2-2b). So the single-model call was threshold-sensitive, which is one reason we no longer rely on it. Under the calibrated turn-matched null, where 12 of 13 models are significant at 95%, the conclusion does not depend on singling out a model. We repeated the sweep under that null. The shared space holds at 99% (12 of 13 models above z = 2, all but Qwen3-32B) and weakens at 90%, where only 3–7 dimensions are retained: 5 of the 9 models with a usable null exceed z = 2, and four give a degenerate null at k = 3–4. Qwen2.5-32B is significant at all three thresholds under both nulls. We also swept layers for the two large Qwen models under the original null; z varies strongly across depth (Qwen2.5-32B from −18.8 to 1.8 at earlier layers), and we report it in the appendix.

Q2/W2: Computational cost.

All runs used NVIDIA Quadro RTX 8000 GPUs (48 GB) in bf16, with one GPU for models up to 14B, two for 24–32B, and four for 72B. Each per-model job covers activation extraction (2000 episodes × 2 roles, a batched forward pass over the 4-turn dialogue, saving hidden states at the probed layers and positions) and the full analysis in both configurations:

Model sizeGPUsWall-clock
2–3B (gemma-2-2b, Qwen2.5-3B)10.75–0.9 h
8–9B (Ministral-3-8B, gemma-2-9b)12.1–2.7 h
14B (Ministral-3-14B, Qwen2.5-14B)13.8–4.1 h
24–27B (Mistral-Small, gemma-2-27b, Magistral)25.0–9.5 h
31–32B (Qwen2.5-32B, Qwen3-32B, gemma-4-31B)27.4–7.7 h

The twelve-model sweep in the submission took about 58 h wall-clock, about 101 GPU-hours; extraction dominates. The narrator baseline (three roles, Qwen2.5-32B) took 10.8 h on 2 GPUs. The analysis works in the retained PCA space, so once activations are saved its cost does not grow with model width: for our widest model (Qwen2.5-72B, d = 8192) the centered analysis took 1 h 20 min on 16 CPU cores with 4.6 GB peak memory, and the new null controls take 1–20 minutes per model on 4 CPU cores. The turn-stratified probe ran as separate CPU-only jobs (8 cores, 32 GB): 1–13 minutes per model.

W3: Typos.

Fixed: the duplicated period in §1, and a further pass for punctuation throughout.

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

The paper is studying whether LLM agents can develop a shared internal representation when communicating with one another and propose a framework that separates and agent's misrepresentation into a share frame and its own internal state and its representation of the other agent. The use techniques like PCA to identify the shared subspace from model activations and follow change across different dialogue turns.

Critical strengths
  1. The authors are very reasonably transparent about the limitations and the shortcomings of their empirical findings.
  2. Even though in a way a negative result is shown but I really liked the use of within-episode centering and turn-stratified probing as it essentially uncovered some statistical confounding that turns out to be a useful tool here. Specifically this proved to be important to show why evaluating representations without accounting for the task structure can lead to misleading conclusions.
Critical weaknesses
  1. I believe the main weakness is that paper's attempt essentially ends up being a negative results as only 1 model show evidence of a shared frame whereas all others perform near chance essentially when tested on the turn requiring actual disambiguation.
  2. While conceptual model of the paper is mainly involving the combination of the three components (F, I, O) but the main results are focusing more on identifying F and its residual geometry rather than a recovery of I and O. I don't think that passing the event-count threshold is the same as showing successful recovery of all the latent components.
Questions for authors
  1. Please have experiments that validate the semantic interpretation of F and improve the method for recovery of I and O.
  2. I think the authors need to check theorem 1 to see if under ordinary independent additive gaussian noise it holds still.
  3. Can you please demonstrate that the statistic actually measures grounding rather than right now where is simply a particular form of representation stability?
  4. Lastly, the abstract has so many symbols that are definitely not needed to be there and in fact challenges the readability a bit.
2 comments

We thank the reviewer for pointed questions. The reviewer's Q3 turned out to be the right question, and the new controls answer it; we address the questions in order of how much they bear on soundness.

Q3: Does the statistic measure grounding, or only representational stability?

In this design, neither specifically: it measures input shared between the roles, mostly text. We ran four tests.

  1. Turn-matched null. We re-pair partners across episodes at the same turn, which keeps the turn schedule fixed. The shared space survives in 12 of 13 models (Qwen2.5-32B z = 10.2), so it is episode-specific, not turn structure.
  2. Narrator baseline. A non-participant reading the same transcript aligns with each role more closely than the roles align with each other (z = 53.0 and 38.6 vs. 10.2). Episode-specific shared text is enough to produce the alignment.
  3. Surface-matched null. Re-pairing only within episodes that share item, color and goal verb does not absorb the alignment (Qwen2.5-32B z = 18.9), because the turn text still differs within those cells. A variant that also required a different grounding branch had too few cells to be usable, and we do not report it.
  4. Branch contrast. We compared the residual outside the shared space between resolved and unresolved episodes at each turn, subtracting each episode's pre-divergence baseline and using an episode bootstrap. The pre-divergence turns serve as a placebo and show no effect in most models. At the Director's reply, where grounding is established, the difference is null in all 13 models (every 95% CI includes 0). Differences appear only at the final turn, whose wording differs by branch, and their sign varies by model (positive in 7, negative in 4).

So the reviewer's alternative is right: in a scripted dialogue the measured alignment does not distinguish grounding from shared input. We have revised H1's claim accordingly (see the general response) and added the design that would separate them: paraphrased or translated scripts per role, with the narrator as a control.

Q2: Does Theorem 1 hold under ordinary independent additive Gaussian noise?

Yes, with a correction to how we stated it. The factor model (App. A.1) already includes isotropic noise, independent across agents. The cross-covariance gets no noise contribution, so the Procrustes rotation is unbiased at the population level. But on a frame direction the residual has variance 2σ², not zero. The correct statement is that F is the eigenspace of the disagreement-conditioned residual covariance at the noise floor, and identification needs the divergent directions' excess variance to exceed that floor by a margin, an eigengap or signal-to-noise condition. Our proof sketch omitted it, and Theorem 1 is now stated with it. Noise correlated across agents, such as shared template tokens, is indistinguishable from frame: that is exactly the shared-text effect in Q3.

We checked this on synthetic data with a planted episode-level frame (true r_F = 2), sweeping noise and sample size (5 seeds per cell, turn-matched null):

Noise σ²0.0030.010.030.10.3
Angle F̂ to true F (median, n = 2000)0.0°1.0°37°81°84°
r_F bias0−1+4+5+4
z (turn-matched null)16.28.51.2−6.3−8.5

Recovery is accurate up to σ² = 0.01 and breaks down by σ² = 0.03, as the eigengap condition predicts; past that point r_F is overestimated, so we treat r_F as an upper estimate. The same sweep shows why we replaced the null: with no frame and only shared turn structure, the row-shuffle null gives z ≈ 5.3 (a false positive) while the turn-matched null gives z ≈ 0.6.

Q1: Validating F semantically, and recovering I and O.

We agree that passing the event-count thresholds is not recovering I and O, and the paper no longer suggests it.

  • I/O split. We tested the EM procedure of App. A.3 against ground truth before reporting it. On synthetic data built so that the resolution change lies exactly in the Follower's true self-subspace, the estimator recovers F within 0.1°, but its self-subspace is 38° from the truth and 52° from the other-subspace, barely better than an arbitrary split of the 2-dimensional divergent block, and it leaks into F. On Qwen2.5-32B something real does happen at resolution: the change falls along a single direction outside F (2nd percentile against random directions). But whether that direction is labeled self or other flips with the definition used: the EM definition gives the Follower's self-subspace, as Theorem 2(a) predicts, and a displacement-based definition gives the other-subspace. Resolution events also cluster at the turn where the branches' text diverges. We therefore do not claim the split, and read the mismatch as a reason to revise the self/other account so that it predicts something about turn-to-turn change that does not hinge on a labeling.
  • Semantic content. The turn-stratified probe is at chance at turn 2 in all models, matching behavior. Projecting its decision direction, 43–54% of its weight lies in F̂ for Qwen2.5-32B (78% for a random direction) and 85% for Qwen2.5-72B (62.5%). We report this as descriptive.

W1: Negative result.

The result is now clearer, and we hope more useful. The positive findings are a validated estimator, a shared space recovered in 12 of 13 models and across model families, and three confounds that make naive analyses misleading, two of which the reviewer singled out. The negative finding is that scripted dialogue cannot separate frame from text, and the paper now gives the designs and alignment methods for doing so.

Q4: Symbols in the abstract.

Agreed; the abstract is rewritten without symbols (general response).

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

The motivation behind the paper is the idea there are cases where agent 'behavior' may be aligned but learned representations are not. The paper looks to the LLM activations to measure grounding and propose a grounding estimator. Using a scripted director-follower task, where the model plays both roles, they search for shared frames using their alignment metric. An analysis finds a shared frame is only present in one model, Qwen2.5-32B, and the paper presents a stability analysis on said model. Probing decodes the content cross-turns but is only chance on the turn requiring said information, indicating that grounding information is present but may not be tracked.

Critical strengths

-The paper is motivated by an interesting problem that many sidestep---what if agents' behavior does not align with the others' representation? This is often sidestepped because studying representation is difficult on its own merit and studying the alignment of representation and behavior only increases the complexity. This all to say, the paper is well motivated. Given the complexity of the problem, the authors have worked hard to constrain the their experiments by focusing on within-episode centering (removing variance) and turn-stratified probing. -The authors deeply examine their results, identifying and discounting several false positives. -Authors are upfront about their limitations---this is an ambitious paper tackling a complex subject but via the limitations keep it scoped. -Methodologically, the paper has a lot to contribute wrt the turn-stratified probe and the decoding of content at the turn-level, and comparing against a chance baseline. The identification and correction of a naive version of the test is meaningful.

Critical weaknesses

-Since the dialogue itself is scripted and sequenced, outputs (e.g., behavior) do not correspond with representation and is arguably not really being looked at. while this controls behavior, it also is not a true expression of behavior---instead, only representation is captured with the experiments. I believe this would have been better captured with an alternative baseline explicitly targeted at capturing single-model prompt representation. -This makes picking apart behavior (opponent shaping, collusion detection) difficult; however, it would be interesting to examine when such behavior characteristics emerge in the static context and try to examine representation from that perspective. However, arguably such experiments would still be subject to the same scrutiny from a behavior perspective. -Additional details would strengthen claims around identifiability criteria and pooled results. What were the exact margins/constraints behind the scaling bounds experiments? Transparancy of other variables factored into dismisal of gemma-4-31B-it, i.e., k. What were the exact rules for mapping predictions -> event labels? etc.

Questions for authors

-Can you clarify the exact rules/heuristics used to generate event labels, dismiss models from additional experiments? Are there additional variables that were factored in but not discussed in the section due to space?

2 comments

We thank the reviewer for the careful reading and for recognizing the turn-stratified probe and the correction of the naive repeated-measures null as methodological contributions. The reviewer's suggested single-model baseline turned out to be the most informative control we ran; please see also the general response.

Q1: Exact rules for event labels, model exclusion, and any unreported variables.

Event labels. At every turn t, both roles receive the dialogue so far and the same forced-choice query, "Does the Follower know which item is meant? (True/False)." The prediction is the argmax over True/False logits. Each role's answer is graded against the ground-truth state of the Follower at that turn. The label is then:

  • agree: both roles answer correctly, and the previous turn was not a disagreement;
  • resolve: both roles answer correctly, and the previous turn was a disagreement;
  • disagree: at least one role answers incorrectly.

No other variable enters the label. In particular, the scripted turn structure (the generator's "expected event type") is logged but never used, which is the exogeneity requirement of Theorem 2(c). The rule never emits the null class ∅ listed in §3; we now say so and remove ∅ from the description. The rule is detect_event_exogenous in utils/subspace_id.py in the supplementary code.

Identifiability thresholds. The thresholds use the scaling form of App. A.8 with target resolution η = 0.1 and confidence α = 0.05: n_needed(C1) = n_needed(C2) = ⌈(r_eff + log 1/α)/η²⌉ and n_needed(C3) = ⌈(⌊r_eff/2⌋ + log 1/α)/η²⌉, where r_eff is the number of PCA components retaining 95% of variance. In preparing this response we found two problems, which we now disclose and correct.

(a) The separation margins γ and ζ were set to 1, not estimated. Since they are at most 1, the thresholds are lower bounds: passing them is necessary, not sufficient. We now estimate γ from the disagreement-conditioned canonical correlations. With estimated γ̂, Qwen2.5-32B (γ̂ = 0.63) needs 3005 disagreement events and has 2063, so it fails C2; only Qwen2.5-14B and Qwen2.5-3B pass C2 with estimated margins. Our estimate of ζ depends on how the self/other split is defined, which we could not validate (see our response to qJff), so we do not claim C3 with estimated margins.

(b) Table 2's C1–C3 columns were computed at the pooled r_eff, while the pre-specified rule uses the centered one. The revised Table 2 reports both, with assumed and estimated margins.

Model exclusion and the confirming rule. The rule was fixed before the sweep: a model confirms H1 if (i) its centered alignment residual is more concentrated than the permutation null, z > 2, and (ii) its event counts pass C1–C3. No other variable was used. As submitted, (i) used a row-shuffle null. We now know that null is miscalibrated (general response, point 1), so we report the rule's outcome under both nulls. Under the calibrated turn-matched null, (i) holds for 12 of 13 models, so it no longer singles out Qwen2.5-32B, and the narrator baseline below shows why. gemma-4-31B's identifiability failure is largely mechanical: its effective dimension is unusually large (r_eff = 61), and the thresholds scale with it.

W1/W2: Scripted dialogue; a single-model baseline is needed. We agree this is the central design trade-off. The script fixes each turn's content so that differences between roles can be attributed to perspective rather than to divergent generated text. The reviewer's concern is that, because one model plays both roles over nearly identical text, alignment may reflect "the same model reading the same prompt." We ran two controls, and together they confirm the concern.

  1. Narrator baseline. A third, non-participant role reads the identical transcript as an uninformed observer and receives the same query. If the Director–Follower alignment were only shared-prompt structure, a narrator should align at least as well. It does, and more: under the turn-matched null, Director–Narrator z = 53.0 and Follower–Narrator z = 38.6, against Director–Follower z = 10.2 (Qwen2.5-32B; the narrator pass reproduces the Director and Follower activations exactly). The two roles without the private information align most closely.
  2. Cross-model pairing. All models ran on the same 2000 episodes, so we paired one model's Director with another model's Follower (25 ordered pairs across Qwen2.5-14B/32B/72B, Mistral-Small-24B and gemma-2-27b). Shared weights cannot explain alignment here, yet 16 of the 20 cross-model pairs are significant under the turn-matched null, up to z = 104 (Mistral-Small-24B × gemma-2-27b), above every self-pairing (7–13).

We interpret these as follows: The estimator recovers a shared space that is robust across models and families, which is a positive and, we think, interesting finding about convergent representations of a common input. But in this design that space is dominated by the shared text, so it is not evidence of a grounded frame. A surface-matched null (re-pairing within the same item, color and goal verb) does not remove it, because such cells still differ in the turn text.

The fix is in the design, and we describe it in the new §7: give the roles paraphrased or translated versions of the dialogue so that they share the situation but not the tokens, keep the narrator as a standing control, and move to free generation, where the model produces its own moves, as the next step toward the opponent-shaping and collusion settings the reviewer mentions.

Comments on the paper

2 comments

How we have revised the paper.

  • The claim becomes: A validated estimator recovers a shared low-dimensional space between dialogue roles across thirteen models and across model families; three confounds (pooling across episodes, a turn-blind permutation null, and shared input text) make naive readings of it misleading; and in scripted dialogue the shared input dominates. The single-model positive call and the capability-threshold discussion are withdrawn.
  • Contributions are split into proposed (the frame/self/other factorization with grounding registry; H2), formalized (Theorems 1–2 and the sample-complexity bound, now with an explicit signal-to-noise condition and estimated rather than assumed margins), and empirically tested (the shared space, the three confounds, the repeated-measures result and the turn-stratified probe).
  • Transparency. Exact event-labeling rule, thresholds and confirming rule, including two simplifications we had not stated (margins set to 1; Table 2's conditions computed in the wrong configuration).
  • New controls and validation: synthetic noise × sample-size sweep, turn-matched and surface-matched nulls, resolved-versus-unresolved branch contrast, cross-model pairing, narrator baseline, PCA-threshold and layer sensitivity, and compute cost.
  • Next steps (new §7). (i) Decouple text from frame: give the two roles paraphrased or translated versions of the same dialogue, vary the script templates, and keep the narrator as a standing control, so that what the roles share is the situation rather than the tokens. (ii) Alignment methods robust to a sparse frame signal: once shared text is reduced, the frame may occupy few directions against a dense background, which dense Procrustes on a 95%-variance PCA basis handles poorly; we will test sparse or regularized alignment and alignment in sparse-autoencoder feature space, always calibrated with the turn-matched null and narrator control. (iii) Revise the self/other account so that it makes a prediction about turn-to-turn change that does not depend on an arbitrary labeling, before estimating it again.

We think the paper is more useful in this form: the controls the reviewers asked for turned an ambiguous single-model positive into a clear, replicated finding about what the measured space is, and into a concrete design for measuring the frame itself.

We thank all three reviewers for their careful and constructive reviews. Every review highlights the same methodological contribution: within-episode centering and turn-stratified probing expose confounds under which pooled analyses look positive. Every review also asks, in different words, the same substantive question: Is the shared structure we measure a grounded frame, or something simpler that two perspectives on the same dialogue would share anyway? We ran the controls the reviewers suggested. They give a clear answer and it changes the paper's main empirical claim.

What the new controls show.

  1. A shared space is recovered robustly across models. We replaced the row-shuffle permutation null with a turn-matched null (partners re-paired across episodes at the same turn). On synthetic data with known ground truth, the row-shuffle null is miscalibrated: it reports a "frame" at z ≈ 5 when the only shared structure is the turn schedule, and z ≈ −4 when there is a frame but the turn structure overlaps it. The turn-matched null is calibrated (median z ≈ 0 with no frame) and recovers a planted frame to within 1° at low noise. Under the calibrated null, the two roles share a low-dimensional aligned space in 12 of 13 models (11 clearing the Bonferroni threshold), and the alignment holds across models: 16 of 20 pairings of one model's Director with a different model's Follower are significant, several more strongly than any model with itself.

  2. In this design, the shared space is dominated by shared text. A non-participant narrator, reading the same transcript as an uninformed observer and asked the same question, aligns with each role more tightly than the two roles align with each other (Qwen2.5-32B, turn-matched z: Director–Narrator 53.0, Follower–Narrator 38.6, Director–Follower 10.2). The tightest pair is the two roles that lack the private information. In a scripted dialogue where both roles condition on nearly identical text, the frame cannot be separated from the text. This is a limitation of the intervention's design rather than of the estimator, but it means we can no longer claim that a grounded frame is detected in Qwen2.5-32B specifically. That model's distinctive result came from the miscalibrated null; under the calibrated one it is one of twelve.

  3. Something happens at the turns, but not what the theory predicts. In Qwen2.5-32B the change at resolution falls along a single direction outside the shared space (2nd percentile against random directions), and in both large Qwen models 71–77% of the turn-to-turn shift lies outside it. But the theory's assignment of that change to the partner's self-subspace (Theorem 2a) is not supported: whether the direction is labeled "self" or "other" flips with the definition used, and our split estimator fails to recover a known self/other split on synthetic data. We interpret this as a reason to revise the self/other part of the theory before pursuing it further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.