- 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
- 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).
- 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.
- 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
- 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.
- 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.
- 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
- 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?
- 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.
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:
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.