- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates how prompt budget, retained memory capacity, and model expressivity jointly constrain the outputs accessible to a frozen language model. The authors introduce a causal-cut formulation and derive theoretical accessibility bounds based on memory capacity, state-update dynamics, and reachable-state geometry, with further specialization to Gated Linear Attention (GLA). Experiments on pretrained GLA models, together with supporting analyses of Sliding-Window Attention and Mamba, examine how memory size and prompt budget affect exact target-sequence recovery.
Critical strengths
-
Interesting theoretical perspective. The paper provides a unified perspective on the relationship between prompt information, retained memory, and model expressivity. The theoretical analysis of memory dynamics and reachable-state geometry offers useful insights into the limitations of memory mechanisms.
-
Controlled experimental investigation. The experiments systematically examine how different memory capacities and prompt budgets affect output accessibility. The additional comparisons involving Sliding-Window Attention and Mamba provide complementary evidence across different memory architectures.
Critical weaknesses
-
Limited validation of the theoretical bounds. The theoretical analysis relies on assumptions about memory-update stability, contraction, and reachable-state geometry. However, the experiments do not sufficiently establish whether these assumptions hold in the evaluated pretrained models or how tight the derived bounds are in practice. Consequently, the connection between the theoretical results and the observed model behavior remains largely qualitative.
-
Insufficient experimental evidence for the central joint-limit claim. The experiments mainly demonstrate that increasing retained memory can improve exact sequence recovery under optimized continuous prompts. However, they do not directly establish when model expressivity, rather than memory capacity or prompt budget, becomes the limiting factor. Furthermore, the primary GLA experiments use relatively small sets of synthetic target sequences, and the findings are not validated on practical language understanding or reasoning tasks. This limits the broader applicability of the conclusions.
Questions for authors
-
Can the authors provide empirical evidence that the assumptions underlying the stability-based and geometry-sensitive bounds hold for the evaluated pretrained GLA models? How tight are these theoretical bounds relative to the observed accessibility frontiers?
-
Can the authors demonstrate an experimental setting in which model expressivity, rather than retained memory capacity, becomes the dominant bottleneck? This would provide more direct validation of the proposed joint-limit framework. Additionally, how well do the findings extend to practical language modeling or reasoning tasks beyond optimized target-sequence reconstruction?
7 comments
I think the theoretical part of this paper is very good.
The analysis is solid, and the main paper and appendix provide detailed derivations. I particularly appreciate the analysis of memory-update dynamics and reachable-state geometry.
Thank you for adding the experiments in the rebuttal.
The additional GLA experiments help connect the theoretical analysis to pretrained models, and the SQuAD results extend the evaluation beyond synthetic sequence recovery.
Overall, I still think the experimental part needs to be strengthened.
In the new Qwen3.5 experiment, increasing the prompt length from 25K to 30K did not improve recovery under the fixed optimization budget. It is still unclear whether this is due to a limit in model expressivity or a difficulty in optimizing the longer prompt. Stronger evidence on this point would make the paper’s main conclusion about the joint limits more convincing.
I will reconsider rating.
Thank you for your thoughtful follow-up and positive comments! We agree that the previous experiment could not clearly distinguish model limitations from optimization difficulty.
To address this, we extended our Qwen3.5 single-head, single-active-layer experiment to 40K prompt tokens and 40K optimization steps, using five paired targets. We measure the output-length frontier at which at least 3/5 targets can still be recovered exactly. All brackets have been refined to 512-token resolution, with exact recoveries verified through native greedy generation.
| Prompt tokens | 20K steps | 30K steps | 40K steps |
|---|---|---|---|
| 10,000 | 9,728–10,240 | 11,264–11,776 | 11,264–11,776 |
| 20,000 | 12,800–13,312 | 12,800–13,312 | 13,312–13,824 |
| 25,000 | 13,312–13,824 | 13,312–13,824 | 13,824–14,336 |
| 30,000 | 13,312–13,824 | 15,360–15,872 | 15,360–15,872 |
| 35,000 | 15,360–15,872 | 15,360–15,872 | 15,360–15,872 |
| 40,000 | 14,848–15,360 | 15,360–15,872 | 15,360–15,872 |
We observe two important things.
First, additional optimization initially improves recovery, but the improvements begin to level off. For instance, with a 30K-token prompt, the recovery frontier increases from approximately 13.5K to 15.5K output tokens when increasing optimization from 20K to 30K steps. However, increasing optimization further from 30K to 40K steps produces no additional shift for the 30K, 35K, and 40K prompts. The endpoint recovery counts also remain unchanged. This gives us greater confidence that the observed flattening is not simply due to insufficient optimization at earlier budgets.
Second, the frontier shows clear signs of plateauing with increasing prompt length. At both 30K and 40K optimization steps, all three largest prompt lengths (30K, 35K, and 40K) yield the same 15,360–15,872 recovery bracket, despite the additional prompt capacity.
We recognize that further optimization could potentially still improve recovery, and our available compute during the rebuttal limits how extensively we can investigate this. Nevertheless, these experiments strengthen the evidence for diminishing returns under restricted model capacity, with the plateau persisting across both larger prompt lengths and optimization budgets.
We will include this additional experiment and its analysis in the final version of our paper.
We appreciate the reviewer raising this point, and hope these additional experiments address your remaining concerns.
Thank you for recognizing our joint perspective. We are glad you found it interesting. We agree deeper empirical validation would be helpful, but we were also limited by available compute.
First, we want to clarify the central goal. We want to establish joint limits and demonstrate why prompt budget, retained memory, and fixed-model capacity must be considered together. Identifying the precise trade-off in a pretrained model is important future work.
In our view, the main practical value is to recognize memory and model capacity as resources to balance when designing language models. People already do this implicitly. Most models have context-window limits reflecting practical constraints, even if these are not absolute expressivity ceilings. Meanwhile, recurrent and hybrid sliding-window or linear-attention models restrict retained memory for efficiency while preserving useful performance. The intuition behind our paper already exists and has been arrived at organically in practice, but these two perspectives have largely been treated separately. We formalize this intuition, unifying these two perspectives. We hope this motivates researchers to explicitly study the balance between memory and model capacity.
Q1. Do stability and geometry assumptions hold in pretrained GLA? How tight are the bounds?
We agree that we did not directly test how well the theoretical assumptions describe pretrained GLA. We have conducted several additional experiments.
Stability: We tested whether small differences in a prompt gradually disappear from the model's memory. We took two slightly different prompts, then fed the model identical subsequent tokens and measured how far apart their memory states remained. After 1,024 tokens, the difference had decreased by about 25% on average, with 191–192 out of 192 trials showing a decrease. This provides evidence of memory contraction under these conditions.
Geometry: We tested whether the memory states occupy a restricted region rather than filling the entire available space. We fitted an ellipsoid using separate calibration data and found that it contained all 48 held-out prompt-end states. Furthermore, preserving just 64 informative directions reconstructed these states with 3.57% average relative error, compared with 69.36% using random directions. This provides empirical evidence that GLA memory has a highly nonuniform structure.
Tightness: We do not claim that our bounds are tight for pretrained GLA, but rather that incorporating memory dynamics and reachable-state geometry allows us to derive tighter bounds than more general capacity arguments. We believe obtaining tighter, model-specific bounds is an important direction for future work. In particular, introducing additional assumptions about data distributions and using information-theoretic techniques such as mutual information could provide more refined, distribution-dependent bounds.
Q2. Can model capacity become limiting? Do the results extend beyond cramming?
We agree this deserves more direct tests. Given available compute and rebuttal time, we completed a deliberately restricted Qwen3.5 single-head, single-active-layer ablation, replacing the remaining blocks with identities. Under a fixed 10,000-step optimization budget, the discovered exact-recovery frontiers (four targets per prompt length) were:
The frontier initially grows but shows clear signs of plateauing, with no gain from 25K to 30K prompt tokens.
Beyond synthetic cramming, we added a 500-example SQuAD-completion experiment on frozen Qwen3-1.7B. We vary the sliding-window size in all 28 layers during ingestion and generation. Accuracy measures answer containment in generated completions, not standard SQuAD exact match.
These are relatively short-context examples: 99% of inputs are at most 512 tokens, and all input-plus-output sequences fit within 1,024 tokens. This shows how sliding window size directly translates to QA performance.
In response, we will make the following revisions:
We hope this appropriately addresses your concerns. Please let us know if you have any further questions