acceptodds
Under review as a conference paper at ICLR 2027

Anchor to the Receiver: Path-Dependent Communication of Quantized Updates

Abstract

Quantized models deployed on many devices are adapted and re-delivered as new data arrives. When each receiver stores a single fixed-grid integer model, every update is a patch against the codes it already holds, and every code change is paid for when it is sent, including changes that a later update undoes. The cost of a stream of updates therefore depends on the path of delivered models, not only on its endpoint. We show that the reference a weight regularizer remembers determines this path. Anchoring to the sender's current quantization bin locks in boundary crossings before they are delivered, whereas anchoring to the installed model pulls delivered changes back. We propose QCTR, which anchors continuous sender weights to the receiver's last-delivered codes, advances the anchor only when a patch has been applied, and keeps the sender's within-cell residual. Changing only the reference lowers traffic at every regularization strength we test, by up to 59%. On a Llama-3.2-1B classification stream, QCTR reaches 87% accuracy with 3.5 less traffic than the strongest alternative. Over ten deliveries it sends 40% less than either alternative reference at similar accuracy and 10.8 less than straight-through fine-tuning. The advantage transfers to Llama-3.2-3B without retuning, holds on a generative stream, and composes with replay and low-rank updates.

Primary area:
foundation or frontier models, including LLMs
Keywords:
quantization, QAT, training, continual learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Tui5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how to reduce the cost when update new model to mobile devices/receivers. When the server/sender trains a new model, sending the updated model to the receiver will impose a communication cost. To minimize such cost, regularization to impose sparse updates is one solution to reduce the transmission cost. This paper studies how to select the anchor for regularization and shows that the proposed QCTR achieves lower communication cost under a multi-round update setup.

Critical strengths
  • The paper targets an interesting question
  • The proposed anchor effectively reduces the cost over multiple rounds.
Critical weaknesses
  • The readability of this paper needs to be improved. For example, the introduction section provides no background. When one starts reading from the first section, terms like “round”, “adaptation”, “receiver”, “sequence of models”, and “in-place updates” cannot be understood with the limited context. Readability is fundamental to a paper. Further revision is needed.

  • The problem is not well motivated. Who cares about this problem? Why is this an issue? These are not even properly provided in the introduction. Specifically, the proposed anchor effectively reduces the cost by several MB. In practice, patches of several MB is common for software apps. Is it really worth the potential performance risk? If it’s an important issue, what’s the current solution? There’s no reference to the closest prior work with a similar setup to motivate the problem. This is also reflected by the fact that none of the baselines are designed for this specific issue. Though it’s not wrong to address a new problem, if that is the case, more discussion needs to be provided to motivate the issue.

  • The survey is not complete. For example, similar prior work [a] studying multi-round model transfer cost is not cited or discussed. This makes the claim of defining the question as a contribution less convincing.

[a] Learning a Neural Diff for Speech Models. Interspeech 2021

Questions for authors
  • As mentioned in the weaknesses, the current writing does not provide enough context for the terms mentioned. Those terms can mean different things in the context of quantization and federated learning literature. Without the context of the target task, it’s hard for the reader to understand the exact meaning.
  • It seems that the proposed anchor achieved better performance than the unregularized counterpart. This needs to be discussed.
Reviewer_oYyNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to reduce traffic when repeatedly updating a fixed-grid quantized model. Each patch is relative to the receiver’s current codes, so repeated changes must be sent again. QCTR pulls continuous sender weights toward the last-delivered quantized state, updates that reference only after delivery, and retains the unsent residual. Experiments compare accuracy and total patch traffic.

Critical strengths
  1. The method addresses communication cost accumulated along an update path. A temporary code crossing followed by a reversal can create patches that must be transmitted. A current-bin anchor can reinforce a temporary change, while an initial-state anchor can undo an already delivered one. QCTR instead anchors to the receiver's last-delivered codes and advances that anchor only after successful delivery. Proposition 1, Eq. (4), and Figs. 2–3 support this explanation. The main contribution is the reference state and its update timing; L2 regularization and error feedback have precedents.
  2. The traffic saving is clear and reasonably controlled. Over ten deliveries, QCTR uses 9.49 MB at 89.64% accuracy, current-bin uses 15.81 MB at 89.38%, and initial-state uses 15.72 MB at 89.51%. Traffic falls by approximately 40% at similar mean accuracy. Code-patch methods share a lossless encoder, with matched-strength tests, low-learning-rate STE, and extended parameter sweeps. These controls reduce the chance that encoding or tuning alone explains the gain.
  3. Ablations and additional tests support the result. Clearing the sender residual after each delivery loses 1.42 accuracy points while saving only 0.62 MB. QCTR plus replay reaches 88.52% at 8.15 MB, close to replay alone at 88.57% and 52.61 MB. Tests also cover several seeds, task orders, hyperparameter transfer to 3B, and a generative stream.
Critical weaknesses
  1. Weight distance does not directly measure transmitted bytes. Proposition 1 gives each code change a penalty proportional to its squared group step. Two changes with similar transmission costs can therefore receive different penalties because their cells have different widths. Appendix A.2 proposes step normalization but does not test it. The cited Neural CSI Compression Fine-Tuning penalizes update rate directly. The results support the receiver reference, but do not establish why the chosen L2 penalty is preferable to an objective that more directly measures communication.
  2. The demonstrated benefit depends on a narrow deployment setting. Tests use 1B/3B models from one Llama family, a fixed INT4 grid, and receivers that obtain every patch at synchronized task boundaries. Skipped or asynchronous updates leave receivers in different states that one anchor cannot match simultaneously. Section 7 acknowledges this, but cross-version patch costs and multiple anchors are not measured. The million-device saving therefore extrapolates the synchronized setting and does not yet establish the benefit for asynchronous deployment.
  3. The reported saving ratios use different comparisons. The 3.5× saving at 87% accuracy comes from interpolating mean frontiers between tested configurations. Primary policies selected by validation accuracy save 11% versus current-bin. The generative 49% saving includes different regularization strengths; at matched strength, it is 18% (Sec. 5.5). The paper distinguishes these comparisons, and each number should retain its conditions. Also, an accuracy interval containing zero does not establish equal quality. The fresh-1B interval, −4.47 to +2.92 points, still permits a meaningful decline.
Questions for authors
  1. Why use the chosen L2 penalty? With the receiver reference, grid, and search budget fixed, can it be compared with Appendix A.2 normalization and code-aware sparsity or rate objectives? Please report quality, changed-code counts, and actual bytes. This would distinguish the benefit of the reference from that of the penalty form.
  2. How much traffic is saved when devices skip versions? Can existing trajectories compare replaying every patch with composing cross-version patches? Please explain how the training anchor is chosen when receivers hold different versions, and report total bytes under one skipped-version distribution.
  3. What saving is reliable after choosing a policy in advance? Can a quality tolerance and byte budget be fixed first, with validation-only selection followed by tests on fresh seeds? Please separate frontier interpolation, matched-strength tests, and independent selection. Quality-difference intervals should show whether the selected tolerance is met.
Reviewer_sgg8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the communication cost of repeatedly updating a deployed quantized model. The authors propose to regularize weights toward the receiver’s last-delivered quantized state, then update this reference after delivery. The analysis distinguishes this reference from alternative anchoring procedures. Experiments show lower cumulative traffic at similar measured accuracy across different tasks.

Critical strengths
  • Reducing the cumulative cost of updating deployed quantized models is an under-explored problem.
  • The regularization's effect on communication is presented clearly and motivated well with respect to the problem formulation.
  • The experiments support the proposed design rule through reference comparisons, residual ablations, and tests with replay and low-rank updates.
Critical weaknesses
  • The methodological novelty appears limited. The main change is the reference used by a regularizer; regularization is already well-established in literature.
  • The quantization grid is calibrated once and remains fixed. This restriction is central to both the analysis and the experiments, but its effect on the accuracy–traffic tradeoff is unclear. Allowing scale adaptation could improve reconstruction while introducing additional communication costs.
  • The method is well motivated under the stated problem, but the practical need for frequent updates is unclear. Which deployment settings require delivery after every task, and when would a few MB per update become a meaningful burden? A concrete use case with realistic update frequency and communication constraints would strengthen the motivation.
  • The headline result of 3.5 MB at 87% accuracy is interpolated rather than directly measured.
  • The abstract uses QCTR without expanding it, and “receiver” appears in the introduction before its role is clearly defined.
Questions for authors
  • What is the main technical advancement beyond applying a penalty relative to the last-delivered quantized model? How should readers distinguish this contribution from existing methods?
  • Why is the initial quantization grid kept fixed? How would adapting scales affect accuracy, patch size, and receiver anchoring?
  • Why deliver an update after every task? Accumulating more training data between deliveries, or updating only when the data distribution changes, could reduce traffic more simply. A comparison with these policies would help establish QCTR’s practical value.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.