- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes a control-theoretic interpretation of Transformer attention, treating queries as active sensing actions, attention weights as filtering gains, and residual representations as belief states. It introduces Iterated Active Attention (IAA), which recurrently updates queries against fixed keys and values with an update-norm stopping criterion, and a reversible residual architecture claimed to provide Hamiltonian, symplectic dynamics with activation storage independent of depth. It also develops interventions and diagnostics for attention sinks, KV compression, and perturbation sensitivity. Validation includes synthetic multi-hop associative recall, token-level stopping profiles on WikiText-103, activation-memory measurements, and diagnostics on Qwen2.5 models ranging from 0.5B to 7B parameters. The reported results suggest that recurrent querying can improve multi-hop retrieval and that some proposed diagnostics are informative; however, they do not establish several of the stronger mathematical equivalence, convergence, and explanatory claims.
Critical strengths
-
IAA reuses fixed keys and values for recurrent querying, with the reported two-hop recall improvement providing a useful proof of concept.
-
The foundation-model experiments are informative, showing both where the proposed methods help and where they fail.
Critical weaknesses
-
The manuscript appears to depart from the prescribed ICLR formatting: the text font is not standard, the title, abstract heading, and section headings use ordinary mixed-case typography rather than the template’s small caps, and table captions appear below the tables.
-
The central mathematical claims are not established for the architecture actually specified. Proposition 1 does not provide a sufficiently specified latent-state and observation model to derive ordinary attention as an exact Bayesian filter, while Proposition 2 assumes natural-gradient dynamics without showing that the learned update , followed by RMSNorm, implements those dynamics. Equations (5)-(8) establish reversible additive coupling, but generic attention and FFN functions need not have the Jacobian symmetry required for canonical symplecticity. Similarly, the row-normalization identity in Eq. (9) supplies no derivation of a Bode sensitivity integral or necessity of an initial-token sink. These gaps affect the paper’s main contribution and require substantive mathematical justification or a substantial narrowing of its claims.
-
The experiments do not isolate a contribution beyond established recurrent attention and reversible architectures, and a central result is internally inconsistent. Multi-hop querying with shared memory and parameters already appears in [1] , adaptive recurrence in [2], and depth-independent activation storage in [3] and [4]. Moreover, Eq. (10) reduces at to the zero-sink operator explicitly evaluated in [5]. Table 2 compares IAA against a baseline with fewer parameters and fewer sequential attention operations, without the matched recurrent and sufficiently deep baselines needed to attribute the gain to the proposed mechanism. Figure 3 appears to show success concentrated at , followed by collapse for , contradicting the caption and text claiming success for and leaving the claimed convergence and adaptive-computation benefits unresolved.
[1] Sukhbaatar et al. End-To-End Memory Networks. NeurIPS 2015.
[2] Dehghani et al. Universal Transformers. ICLR 2019.
[3] Gomez et al. The Reversible Residual Network: Backpropagation Without Storing Activations. NeurIPS 2017
[4] Kitaev et al. Reformer: The Efficient Transformer. ICLR 2020
[5] Xiao et al. Efficient Streaming Language Models with Attention Sinks. ICLR 2024.
Questions for authors
Please see Critical Weaknesses.
4 comments
Brief Follow-Up on Formatting, Soundness Fixes, and Baselines
Dear Reviewer xjgj,
As the discussion window enters its final days (closing Nov 18), we wanted to briefly check whether our revised manuscript and response resolved the formatting, soundness, and baseline concerns in your review. In particular, we:
- Restored standard ICLR Small Caps typography and placed all table captions above tables across the main text and appendix.
- Fixed the ungated script bug behind the drop, achieving 100.0% recall across all and on static KV states against five matched baselines—including Ungated Memory Networks (which collapse at to 34.4-35.6%) and Universal Transformers (Table 1, Fig. 2, App. Table 8)—while crediting prior work [1]-[5] throughout.
- Replaced circular/overstated claims with self-contained proofs for Propositions 1-3, unit-determinant shear invertibility (), and row-stochastic vs. spectral norm bounds (Appendices A-C).
Please let us know if anything remains unclear before Nov 18. Thank you again for your thorough review!
I thank the authors for the reply. However, my concern about the applicability of the theoretical guarantees remains. The reply did not explain clearly how the contraction and entropy descent conditions hold in evaluated models. I therefore maintain my score
How Contraction and Entropy Descent Hold in Our Evaluated Models
We thank Reviewer xjgj for following up. Your point is well taken: our previous reply stated the algebraic inequalities in Proposition 1 without clearly explaining when and how those two conditions hold in standard unconstrained training (where we do not clamp during optimization).
In our trained models, query updates operate in two distinct regimes: intermediate hops (, moving from one bridge key to the next) vs. final-answer steps (, once the target value is retrieved). Both conditions in Proposition 1 govern this convergence regime, as measured directly on our evaluated MQAR models (run_mqar_ablations.py) below:
1. How the Contraction Condition (, Eq. 1) Holds in Evaluated Models
During intermediate hops (), the update map should not be globally contractive: if held everywhere from , would stay near instead of jumping across distinct bridge keys in context (indeed, trained models learn ).
Once the target value is reached (), however, Proposition 1 holds locally without weight clamping because the two learned terms in Eq. (1), , vanish via gate closure and peaked softmax:
- The sigmoid gate learns to close on terminal answers ( drops , from ), and because , gate sensitivity shrinks simultaneously.
- Softmax concentrates onto the target key, driving the local attention Jacobian toward zero while near the stationary query.
- Consequently, , reducing the effective gain toward : empirically, every consecutive step ratio stays strictly below () and relative updates shrink (), triggering Step 4's early halt (as also seen on WikiText-103, where 100% of punctuation and 96.7% of determiners halt at ).
2. How the Entropy-Descent Condition () Holds in Evaluated Models
By Appendix A (Eq. 27), points toward keys that already carry high attention weight. Thus:
- During hop switches (): Because intentionally shifts away from the current bridge key toward a new key, attention entropy briefly rises (, with only 44.1% of test examples satisfying ).
- Once the target key is located (): Cross-entropy loss trains to reinforce the target key, so holds on 80.5-82.0% (), 97.7-98.0% (), and 99.6-100.0% () of test examples. This sharpens attention entropy down to (100.0% accuracy), whereas an Ungated Memory Network () keeps overwriting the query, dispersing attention back toward uniform entropy ( vs. uniform, 30.1% accuracy):
| Step (, ) | Mean Gate | Rel. Step | Step Ratio | Test Examples w/ | IAA Entropy (Acc.) | Ungated Entropy (Acc.) |
|---|---|---|---|---|---|---|
| (Hop 1) | 0.265 | 1.262 | -- | 44.1% | 0.06 (4.3%) | 2.63 (5.9%) |
| (Hop 2) | 0.074 | 0.390 | 0.318 < 1 | 44.9% | 0.52 (75.8%) | 2.11 (99.6%) |
| (Answer) | 0.035 | 0.135 | 0.453 < 1 | 80.5% | 0.42 (98.4%) | 3.12 (19.1%) |
| (Extra step) | 0.032 | 0.105 | 0.778 < 1 | 82.0% | 0.36 (100.0%) | 2.89 (34.8%) |
| (Extra step) | 0.031 | 0.082 | 0.931 < 1 | 75.0% | 0.32 (100.0%) | 2.95 (30.1%) |
We will add this explicit two-regime explanation and empirical diagnostic table directly to Section 3.1 and Appendix A so the scope of Proposition 1 is immediately clear.
Response to Reviewer xjgj
We thank Reviewer
xjgjfor a thorough review that caught the τ>k script bug and formatting issues, and sharpened our mathematical analysis and prior-work discussion ([1]-[5]).W1: ICLR Formatting and Caption Placement
We restored standard ICLR Times / Small Caps typography across the title, abstract, and headings, and placed all table captions above tables in the main text and appendix.
W2: Mathematical Precision of Propositions 1 and 2, Reversible Coupling, and Sink Logits
W3: Prior Work [1]-[5], Baselines, and Resolving the τ>k Collapse
run_mqar_ablations.pyhad implemented an ungated update (q = self.w_update(v_agg)), omitting Algorithm 1's sigmoid halting gate (src/models/iaa.py). Without gating (as in Memory Networks [1]), once τ=k retrieves leaf vjk, step τ=k+1 overwrites qi with Wupdatevjk=kjk, stepping off the target. Restoring Algorithm 1's gate (gi(τ)→0) locks onto leaves across k∈2,4,6 and L,τ∈1..8:W4: Empirical Covariances, Row-Stochastic Norms, and Block Bound (Sec. 4.2.2-4.2.3, App. C)