acceptodds
Under review as a conference paper at ICLR 2027

Predictive Compression for State–Access Allocation

12 views

Abstract

A prediction head must choose how much local state to retain and how often to read a shared representation. State preserves predictable future answers between reads; new task information can require an earlier read. For periodic reads of fixed-width packets and outputs affine in the held state between reads, an analysis separates these bottlenecks, and controlled synthetic interventions exhibit both. We train one multi-step predictor, compress its future task answers into candidate held states, calibrate their executed risks, and deploy the cheapest passing configuration. Across 12 new synthetic worlds, compilation gives a 20.29× cold-preparation speedup over training a complete receiver grid. A matched follow-up on the same worlds yields a 12.11× paired geometric speedup over cost-ordered early-stop receiver search, at 1.038 times the baseline’s selected normalized state-and-read cost. With 65,536 calibration episodes per task, all 432 allocation requests pass independent risk tests at a target mean squared error of 0.04. A sufficiently trained shared grid can select slightly cheaper budgets on these synthetic tasks, while a simple sample-and-hold baseline is stronger on the tested frozen video features.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
predictive state representations, task-oriented compression, intermittent observations, resource allocation, multi-step prediction, risk calibration

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LTyCOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks how to allocate resources to a prediction head attached to a shared representation: how much persistent state to keep (dd FP32 coordinates), how often to read the shared features (a fixed-width input packet every gg steps). It first gives a theoretical split of the cost into two parts. For a decoder whose output between reads is an affine function of the persistent state, the lower bound on the gap risk is νggap+∑i>dλi(Kg)\nu_g^{\text{gap}} + \sum_{i>d}\lambda_i(K_g). The first term is new information that arrives only after the last read, which can be reduced only by reading earlier. The second term is the predictable spectral tail cut off by the finite state, which can be reduced by adding state. On top of this, the paper proposes a concrete method: train a multi-step predictor of the task's next 16 answers, project its predictions onto a dd-dimensional subspace to obtain a deployable head for each candidate (d,g)(d,g), estimate each candidate's risk upper bound using 65,536 independent calibration episodes and an empirical Bernstein inequality, and select the cheapest combination satisfying U(d,g)≤q=0.04U(d,g)\le q=0.04. This removes the need to train a receiver for every candidate.

On validation, main results come from 12 new synthetic environments. The compiled method's cold preparation is 20.29× faster than a full receiver grid, the selected resource cost is 1.038 times that of the grid, and all 432 price queries pass the independent test; against a cost-ordered early-stop search, the speedup is 12.11×. Interventions on 24 worlds show that adding state helps under low innovation and shortening the read interval helps under high innovation, matching the theory. In comparisons, the method that compresses future answers is 32–40% cheaper than the methods that compress current features (P/H), but it essentially ties with the stronger per-cell shared supervised grid. The authors also honestly report unfavorable results: when the calibration size drops to 16,384, 216/432 queries find no certified cell; independently trained receivers fail on high-harmonic tasks; and on frozen DAVIS video features, simple sample-and-hold achieves a lower MSE (by 0.0052) at no greater cost.

Critical strengths

Strength 1: Paper separates state capacity and read frequency into two bottlenecks. Proposition 1 states that, for a decoder whose output between reads is an affine function of the persistent state, the lower bound on the gap risk is νggap+∑i>dλi(Kg)\nu_g^{\text{gap}} + \sum_{i>d}\lambda_i(K_g). The first term is new information that appears after the last read, which increasing the state can't recover. The second term is the predictable spectral tail cut off by the finite state, which increasing the state can reduce. In the low-innovation world, increasing dd from 1 to 7 reduces MSE by about 0.167; in the high-innovation world, shortening gg from 8 to 2 reduces MSE by about 0.181, while d=7,g=8d=7, g=8 can't get past the information floor. These results come from 24 independent worlds, and the statistical unit is the world.

Strength 2: I was able to verify the key parts through the code the authors provided. I spent some time understanding the whole code, on two tasks I recomputed the projections (UUᵀ differing by about 10−710^{-7}), the calibration upper bounds for all 24 cells (differing by about 10−1110^{-11}), and the risk on 524,288 independent test episodes. I think the results show that the selection rule plus the empirical Bernstein certification is reproducible.

Critical weaknesses

Weakness 1: Paper's biggest speedup is measured against a weak baseline. All of the evidence that the method works comes from a synthetic environment. The target is a combination of a finite number of circular modes, and the output between reads happens to fall within the affine class characterized by Proposition 1. Authors themselves also acknowledge some results fail on high-harmonic tasks. The only real-data experiment is on DAVIS video features, its result is unfavorable to the method: simple sample-and-hold has an MSE 0.0052 lower than F at no greater cost, the q=0.04q=0.04 target was adjusted only after the development videos had been inspected, and the 30 videos cannot certify the population risk. 20.29× is measured against a full grid in which the receivers of all 19 cells are trained, which is a very easy baseline to beat. A more reasonable comparison is early-stop search, for which the speedup falls to 12.11×, and that comparison was in turn specified after the fact on old worlds whose results had already been seen; the more conservative F-then-A path achieves only about 5.6×. In addition, this advantage requires 65,536 calibration episodes, and when this is reduced to 16,384, 216/432 queries find no certified cell.

Weakness 2: The lower bound in Proposition 1 holds only for the ideal affine class. Authors themselves point out that the F, P and H actually executed all carry a final output clip and belong to Fclip-aff\mathcal{F}_{\text{clip-aff}}, which is not in this class. Paper also reports that a correctly clipped nonlinear receiver crosses the affine ideal minimum on seven tasks, showing that this bound is not a restriction on all legal receivers. The budgets selected by the method are certified only by an upper bound and not by a lower bound: at λ=4\lambda=4, the blind lower-bound rule excludes all cheaper cells in only 8 of the 72 tasks in each family, so the high agreement with the independently trained grid is mainly an empirical phenomenon. At the same time, the advantage the paper shows is also quite limited: F is cheaper than P/H, but these two comparators merely apply principal-component compression to the current features and have no task-supervised encoder; and against a stronger comparator (a shared grid trained separately for each cell, S-per-cell), F only ties on 70–71/72 tasks.

Questions for authors
  1. In the only real-data experiment (DAVIS), sample-and-hold achieves a lower MSE at no greater cost. On what kinds of real tasks do you expect compressing future answers to outperform simply holding the current estimate, do you have any real-data results showing this?

  2. What are the side-by-side timings of F, the full grid and early-stop search on the same machine? Also, in the stored results, F-then-A has test results identical to Direct-A; did F-then-A actually retrain an independent receiver?

  3. Proposition 1 holds for the ideal affine class, while the executed F, P and H carry an output clip, and selection relies only on upper-bound certification. What role does the theory play in the method's selection rule, beyond motivation?

2 comments

Thank you for checking the projections, 24 calibration bounds and test risks on two tasks. We added a same-packet temperature comparison and the requested same-machine timing study, and clarified how the bounds guide budget selection.

Forecasting and hold on real data. We added a same-allocation comparison on a previously examined Jena station. Forecast compression (F) with one value and eight-hour reads gives normalized MSE .0507 versus .1451 for holding the latest temperature. At sixteen hours, F improves from .1186 with one value to .0694 with two, meeting the .10 target. A supervised same-packet receiver also shows this exchange. The 17,462 temperatures include read steps, packets contain source-side history, and errors use training-temperature variance (revised §4.2 and Appendix H). In the original DAVIS comparison on 30 held-out videos, hold gave .00519 lower MSE at no greater cost (original §4.5 and revised §4.7).

Same-machine timing. We added a side-by-side comparison in one process on one logical CPU (revised §4.3 and Appendix E). On 24 preselected tasks from 12 previously examined worlds, F averaged 1.439 s/task, the full grid 29.117, and early-stop search 18.902. Timing includes fitting and calibration and excludes the shared setup and test.

Refitting and calibration. We now explain in Appendix E why F-then-A has the same test result as Direct-A at a shared cell. It refits the receiver using Direct-A’s training data, seed and deterministic recipe, then calibrates it on an independent lane. In the original comparison, F and Direct-A each certified all 432 choices with 65,536 independent calibration episodes. At 16,384 episodes, F certified 216/432 and Direct-A 380/432 (original and revised Appendix E).

Theory and budget selection. We clarified the role of each bound in revised §§2–3. The selector chooses the cheapest candidate with a passing calibrated risk upper bound. The affine lower bound rules out a budget when it exceeds the target for current-packet writers whose complete between-read output lies in a fixed affine space of dimension at most the state size. At .04, it excludes two values at sixteen steps in the original example, under its source and numerical conditions. It motivates projecting forecasts onto their leading directions. At read price 4, the original blind lower bound excludes every cheaper cell in 8/72 tasks per source family. The upper-bound rule gives feasible choices within that same evaluated grid (original and revised Appendix F). The main compiler’s gap outputs are unclipped (original and revised §3). F/P/H comparison heads use a final output clip (original §4.3, revised §4.4 and Appendix F in both).

Supervised-grid comparison. In the original comparison, after 38,000 updates, the supervised grid ties F on 70–71 of 72 tasks per comparison and wins the rest (original §4.3, revised §4.4 and Appendix F in both).

Thank you for your detailed rebuttal and for providing the additional experimental results. I have read your response and examined the newly added experiments. However, I still have reservations regarding the reliability of the experimental data in your paper. I am maintaining my current rating.

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

In this work, the authors study a specific scenario in which a small model must make a prediction at every time step but may only look at its input every few g steps, relying in between on a handful of d numbers saved at its last look. Two budgets are traded off: more memory d versus more frequent reads (smaller g), with the cost proxy J = d/8 + λ/g. The question is which combination of memory and look frequency meets an error target at the lowest cost.

Authors train one network to predict the task's next 16 answers. For each candidate (d, g), they compresses that forecast into d numbers using PCA.

Each candidate (d, g) is evaluated on a large held-out set to get a safe upper estimate of its error, and the cheapest one that stays below a pre-determined error threshold is chosen. Authors also provide a short theoretical result explaining the trade-off: some error comes from things in the future that were unknowable at the last look, which only more frequent looks can fix, and some from predictable information that did not fit in the d numbers, which more memory can fix.

Validation is mostly on synthetic 9-number signals built from rotating latent oscillators. The main result is a 20.29x faster setup than training a separate model for every (d, g) cell, at 1.038x the selected cost. In a real-data test (DAVIS 2017 video with frozen ResNet-18 features), a simple "store the last estimate and repeat it" baseline is more accurate than the proposed method at no higher cost.

Critical strengths
  1. Careful experiments and honest reporting: Data used for training, model selection, calibration and final testing is kept separate throughout. Negative results are reported honestly, for example that a simple baseline beats the method on the DAVIS video experiment.

  2. Error between reads comes from two sources: things that happened after the last read, and predictable information that did not fit in memory. Authors demonsrate this and the trade-off between more memory and less look-ups clearly.

Critical weaknesses
  1. The paper is very hard to read. A lot of technical jargon is used instead of writing to a non-expert reader.

  2. The proposed setting is not practical. The cost being optimized (J = d/8 + λ/g) does not measure compute, latency, energy or bandwidth. In the synthetic tasks, the trade-off is between storing a few extra 4-byte numbers and reading a 36-byte input, which is does not represent a realistic bottleneck. On the only real-data test (DAVIS videos), the method performs worse than a sanity-check baseline.

  3. The claimed 20x speedup comes from training one model instead of seperate 19. This is a sensible thing to try first but not surprising.

  4. The synthetic signals are noisy rotating oscillators close to a linear state-space system. Yet neither repeating the last estimate (which wins on video) nor a Kalman-style predictor with intermittent observations is tested. The compression baselines using PCA of the input and of hidden features are unsupervised and weak, which makes the reported 40% savings look larger than they are.

Questions for authors

Please address the weaknesses listed above.

1 comment

Thank you for asking about the practical cost and stronger baselines. The revision shows what the two resource bounds tell a designer and adds recorded-temperature and synthetic comparisons.

What the bounds tell the designer. We revised the opening and §§2–3 to explain when to store more or read sooner. For a fixed packet, the logical cost prices stored state and read frequency. The bounds hold the exogenous source, schedule and error target fixed. A schedule is infeasible if even the best prediction from all permitted reads and side information exceeds the target. For current-packet writers whose complete gap predictions lie in a fixed affine space of dimension at most the state size, our spectral bound quantifies the predictable information lost to limited state. We reuse one forecast to construct candidates and use separate synthetic episodes to certify those whose risk upper bounds meet the target (revised §§2–3).

Recorded temperatures and communication. In the added Jena study, increasing state from one to two values let both forecast compression (F) and supervised reduced-rank regression on the same packet double the read interval from eight to sixteen hours at the .10 normalized-MSE target. Both missed that target with one value at sixteen hours. In a local two-process replay on these previously examined data, the two-value, sixteen-hour F configuration used 36.1% fewer application bytes than the one-value, eight-hour configuration, counting requests and synchronization with identical full packets and source-maintained history (revised §4.2 and Appendix H). A next deployment test is to measure end-to-end latency and source energy under the same packet protocol (Discussion).

Sample-and-hold. We added sample-and-hold comparisons on Jena and in the controlled synthetic study. On Jena at one value and eight-hour reads, F had lower normalized MSE than hold (.0507 vs .1451). In a follow-up on 12 previously examined synthetic worlds, both methods passed the test bound at the same .04 MSE target. F selected lower logical cost and hold achieved lower mean test MSE (revised §§4.2–4.3 and Appendices E and H). The original DAVIS hold comparison retains its advantage of .00519 lower MSE at no greater cost (original §4.5, revised §4.7).

Intermittent state-space baselines. In an added follow-up on 12 synthetic worlds, F, an enhanced linear filter using a task-derived proxy state, and neural reset all pass independent test bounds. At error target .04 and read price 1, F selects lower mean logical cost than the other two. F uses extra projection episodes, and a retrospective shared-example projection preserves its choices and test passes (revised §4.3 and Appendix E).

What forecast reuse saves. In the original same-session study, forecast compilation prepared passing allocations in 1.47 vs 26.04 s/task against the long-trained supervised grid at the same error target. Both yielded 432/432 independently test-passing choices. The grid selected slightly cheaper allocations on average (original §4.1 and revised §4.3, with details in Appendix E). The reported 32.5–40.5% logical-cost savings are against hidden-feature PCA (revised §4.4 and Appendix F). The original paper also reported an early-stop timing follow-up on previously examined worlds that favored compilation (Appendix E). The added same-machine rerun reports all three preparation times together (revised §4.3 and Appendix E).

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

The only thing I was able to make out from the paper is that the central question is how much state/memory information a predictive model needs while remaining accurate.

Critical strengths

None. I simply wasn't able to understand the main claims and experiments in this paper.

Critical weaknesses

Sadly, this paper is just not understandable in its current form and the extremely hard-to-parse language makes it difficult to understand what the (simple) question asked here even is. This is critical to my assessment and unfortunately I strongly lean towards rejecting this paper. The claims are not clearly presented and it's unclear what I should take away from this paper. Examples of weird phrasing:

  • "State can preserve predictable future answers across a read gap; new task information can require an earlier read. A whole-gap affine analysis separates these bottlenecks, and controlled synthetic interventions exhibit both." -- right in the abstract.
  • "The source information floor applies to every legal component when the current packet suffices for the lawful read history. For the zero-mean synthetic sources, a qualified spectral-tail certificate can also exclude an entire whole-gap affine cell. Neither is inferred from one failed network fit." -- from Section 3.

These are just some excerpts. I do not think the work has been properly verified, and the AI Use Statement also leads me to believe that a lot of the work (if not all of it) was simply done by AI:

OpenAI Codex assisted with problem and hypothesis refinement, literature screening, mathematical derivations, experiment design, implementation, analysis and interpretation, figure preparation, and manuscript drafting and editing, and supplementary artifact preparation. [...]
Questions for authors

None that may change my assessment of the paper in its current form.

2 comments

Thank you for pointing out those passages. We rewrote the abstract and clarified the §3 passage on how the two resource bounds inform the budget choice. Section 4.2 now shows those choices in an hourly temperature task.

The task. Can hourly temperature predictions meet the same target with one extra stored value and half as many station reads? In our added comparison on a previously examined Jena station, a forecast from a full packet with source-side history gives normalized MSE .0507 with one value at eight hours and .1186 at sixteen. Two values give .0694 at sixteen, below the .10 target. All means use the same 17,462 observed temperatures, include read steps and are normalized by training-temperature variance (revised abstract, §4.2 and Appendix H).

The two resource limits. State can retain changes that can be predicted from the last read. Changes that are unpredictable from the available reads and side information may require an earlier read. For current-packet writers, the affine bound quantifies predictable variation omitted when the entire between-read output lies in a fixed affine space of dimension at most the state size. Independent episodes calibrate a statistical risk upper bound for each fitted synthetic receiver (§§2–3).

Reproducing the reported numbers. The revised supplement includes real/selfcheck.py, which recomputes Jena mean NMSE from saved error sums and counts. A reviewer also recomputed projections, 24 calibration bounds and test risks on two tasks using the supplied code. The authors led problem formulation and evidence collection and take responsibility for the paper. As disclosed, OpenAI Codex assisted with refinement, derivations, experiments, analysis and writing in various degree.

Thank you for the response. I have read all the authors' responses to the reviewers and at this point, I think I will retain my score.

Comments on the paper

1 comment

With a fixed packet format and error target, a predictor can sometimes read less often by storing more values. Extra state can preserve predictive information from the reads it already makes. A schedule can still leave too much uncertainty even with unlimited state.

We thank the reviewers for asking when storing more at a receiver lets it read less. We revised the abstract and introduction around hourly temperature prediction. The predictor forecasts temperature every hour, reads a fixed-width input packet at intervals, and can store a few values between reads. The packet format, task and error target are fixed in each comparison. For a fixed exogenous source and schedule, what cannot be predicted from the permitted reads and side information sets an error floor even with unlimited state. Finite state can also discard variation that was predictable at the last read. A spectral lower bound quantifies this second loss for current-packet writers whose complete predictions between reads lie in a fixed affine space of dimension at most the state size. In §§2–3 we clarify which budgets each lower bound can rule out. For synthetic tasks, we form candidate receivers from one fitted forecast. We select the cheapest candidate whose risk upper bound, calibrated on independent episodes, meets the target.

In the exchange example retained in Figure 1, two stored values suffice with eight-step reads and four suffice with sixteen-step reads at MSE .04. The affine-class lower bound rules out two values at sixteen steps under the stated source and numerical conditions. Figure 3 now identifies which fitted receivers meet the target in the separate resource interventions. It marks where even the full permitted read history leaves too much uncertainty on high-innovation ideal sources at eight steps (revised §4.1 and Appendix A).

In response to the real-data and communication questions, we added the recorded-temperature comparisons in §4.2 and Appendix H. On previously examined Jena data, adding one stored value let both forecast compression and supervised reduced-rank regression on the same packet double the read interval at the same empirical target of .10 normalized MSE. In a local two-process replay with identical full packets and continued hourly source observations, the longer interval reduced application bytes by 36.1% for forecast compression. A separate check on eight held-out NOAA stations found lower sixteen-hour RMSE with more state at all eight stations and the interval exchange at two under the prespecified 2.76°C target. The temperature comparisons use empirical mean errors. The synthetic risk upper bounds use independent calibration episodes.

In response to the request for side-by-side preparation times, the revised manuscript compares all three methods on the same tasks in one CPU session. Forecast compilation prepared passing budgets faster than full-grid fitting and early-stop search. We also added same-target hold and intermittent state-space comparisons of logical budget cost and achieved error (revised §4.3 and Appendix E).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.