acceptodds
Under review as a conference paper at ICLR 2027

TORQUE: Optimizing What (not) to Quantize Before and After Rotation

Abstract

Random rotations are an effective preprocessing step for quantization: regardless of the input, they make coordinate distributions approximately Gaussian, enabling the use of precomputed optimal codebooks. We introduce TrunQuant, a framework that improves on previous work using random rotations by jointly optimizing how many and which coordinates to preserve at high precision before and after rotation, under a fixed overall expected bit budget. Intuitively, before rotation, preserving large input coordinates at high precision can reduce overall error by preventing the rotation from spreading their values across many coordinates. Likewise, after rotation, preserving a small fraction of the largest-magnitude coordinates at high precision allows the remaining values to be quantized more accurately using codebooks optimized offline for the resulting truncated Gaussian distribution. We demonstrate TrunQuant's effectiveness across various biased and unbiased quantizers, including both scalar and vector methods, and show that it improves worst-case accuracy and applications such as KV-cache compression and approximate nearest-neighbor search.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
quantization, random rotations, vector quantization, model compression, approximate nearest-neighbor search

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5CtzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TORQUE allocates an expected bit budget jointly across high-precision retention before rotation, high-precision retention after rotation, and quantization of the remaining values. It retains the largest input coordinates, normalizes and rotates the residual without reducing its dimension, and uses offline codebooks fitted to truncated Gaussian distributions. Under Haar rotation and exact retention, the paper proves that top-k input retention minimizes an NMSE upper bound for each k. A Gaussian surrogate and a largest-feasible-k reduction make parameter selection efficient. Experiments wrap EDEN, RaBitQ and TurboQuant, evaluate captured Qwen2.5-0.5B activations and KV caches, and measure reconstruction and retrieval on GloVe and GIST. Additional results cover HIGGS under the Gaussian model and encoding/decoding time. The main contribution is the joint allocation and its optimization structure; the individual retention mechanisms have prior precedents.

Critical strengths
  1. The joint optimization has a useful and simple structure. Retaining input outliers reduces residual energy while consuming bits that could improve the inlier quantizer. The paper accounts for retained values and indices, and shows that, for fixed k and the stated assumptions, top-k minimizes the residual-energy bound.

Holding the rotation dimension fixed also permits efficient evaluation of the largest feasible k for each codebook candidate (Section 2.3 and Appendix A). This turns a potentially combinatorial design into a practical per-vector choice. The paper appropriately distinguishes the rigorous bound from its Gaussian approximation.

  1. The empirical evidence supports a real reconstruction benefit across several quantizer families. The retrieval evaluation uses the complete million-scale GloVe and GIST databases, and the synthetic pre/post/joint comparison helps explain which input distributions favor each stage (Figures 3, 5, 7 and 8).

Reporting optimization-inclusive timings on multiple platforms is useful for assessing cost. The paper also explicitly acknowledges that lower NMSE does not consistently improve GIST Recall@1.

Critical weaknesses
  1. The practical equal-storage comparisons and unbiased variants cannot yet be adequately verified.

Equations (4)-(5) use a direction-independent rotation model, while the implementations use randomized Hadamard transforms and the KV heads have only 64 coordinates. The exact transform composition and the probability used for feasibility are not specified sufficiently to check the expected budget. A single RHT does not generally have Haar marginals, so the model's feasibility does not automatically imply the implemented expected-rate constraint. Moreover, Section 2 says metadata belongs in the total budget, but the application experiments do not give per-method, per-chunk bit ledgers or measured complete encoded rates; Appendix D explicitly uses a different convention with an additional 160 bits per vector for both methods. The manuscript also does not define the truncated-inlier reconstruction and correction formulas for the named unbiased baselines. These omissions affect the central claim of improved error at equal storage, particularly at small dimensions. They do not establish that the experiments exceed their budgets or that the implementations are biased, but prevent checking those possibilities.

  1. The experiments do not sufficiently establish the added application value of joint retention. Post-rotation retention and truncated-distribution quantization already appear in QUIC-FL, as the paper acknowledges, but the matched pre-only/post-only/joint accuracy ablation is limited to synthetic data and EDEN (Figure 5b).

The real-workload figures do not isolate the incremental gain from the new joint allocation. The LLM experiments measure captured-tensor NMSE for one 0.5B model, without inference using the quantized activations/caches or a downstream quality metric. For KV caches, Appendix C.1 additionally changes both the context count and the retention policy, so it cannot isolate the stage contribution. This matters because GIST demonstrates that NMSE and task quality can diverge, and joint retention, including parameter optimization, increases CUDA encoding time by 48.4-99.7% in the tested setting, with larger overheads in Appendix D. A matched application ablation and an inference-quality result would make the contribution's practical significance much more convincing.

Questions for authors
  1. Budget and rotation clarification: For each application figure, what exactly does the horizontal-axis budget include? Please specify the random transform, the retention probability used by the optimizer, and the bits for norms/corrections, counts, codebook identifiers, indices and padding. Can you report the measured mean complete encoded bits per coordinate at the target budgets, particularly for d = 64, 128 and 256, together with modeled versus observed retention rates? An expected budget is appropriate to the formulation. The question is whether that expectation is met by the actual implementation.

  2. Quantizer clarification: What are the exact reconstruction/debiasing formulas for EDEN, RaBitQ and TurboQuant after truncation, including fractional-bit allocation and the treatment of retained values? Does the TurboQuant comparison retain the original residual-QJL construction or use a different correction? Please give the candidate sets, codebook fitting procedure, accuracy-experiment precision and error-bar definitions, and clarify whether measured NMSE includes all stored-value and scale-rounding errors. These details determine whether the reported comparisons implement the intended baselines.

  3. Evidence needed to resolve the significance concern: Do matched pre-only/post-only/joint results on the same real inputs and complete storage budgets show a material advantage for joint allocation? Is there a result measuring inference quality with the quantized KV cache or activations actually used by the model? Please also explain the relationship of the post-only variant to QUIC-FL. These are targeted checks of the submitted application claims, rather than requests for a new application domain.

Reviewer_W14XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TORQUE, a framework for random-rotation-based quantization that jointly optimizes outlier retention before and after rotation under a shared bit budget. Pre-rotation retention prevents large-magnitude coordinates from spreading their influence during rotation, while post-rotation retention enables accurate quantization using truncated Gaussian codebooks. The authors derive an NMSE upper bound minimized by top-k pre-rotation retention, enabling efficient parameter selection with offline codebooks. Experiments with EDEN, RaBitQ, TurboQuant, and HIGGS demonstrate improved reconstruction accuracy across synthetic, retrieval, and LLM workloads.

Critical strengths
  1. Well-motivated framework. Jointly optimizing pre- and post-rotation outlier retention addresses a key limitation of rotation-based quantization, with explicit modeling of all bit costs.
  2. Clear theory and efficient optimization. The top-k retention optimality proof and efficient parameter-selection algorithm with offline codebooks provide a solid foundation.
  3. Broad evaluation. Tested across multiple quantizers and distributions (synthetic, retrieval, LLM activations/KV-cache) with informative ablations.
Critical weaknesses
  1. Theory-practice gap. The analysis relies on NMSE bounds, ideal Haar rotations, and Gaussian approximations, but lacks validation for finite-dimensional vectors (e.g., 64-dim KV-cache).
  2. Limited downstream evaluation. Focuses on NMSE without LLM perplexity/generation metrics; retrieval improvements are modest.
  3. Significant encoding overhead. 2–8× baseline on CPU and 1.8–4× on Apple GPU, with insufficient cost-benefit justification.
  4. Bit-budget and implementation issues. Constrains expected (not actual) bits, inconsistent metadata treatment, and FP16 vs. theoretical exact retention.
  5. Unclear advantage over simpler alternatives. Joint retention benefits over individual strategies on real workloads remain insufficiently demonstrated.
Questions for authors
  1. How accurately does the theoretical objective predict actual reconstruction error across different dimensions?
  2. How do NMSE improvements translate to LLM perplexity and generation quality?
  3. How are strict bit-budget constraints and metadata overhead handled in practice?
  4. Can you provide comprehensive Pareto comparisons of accuracy, latency, and storage?
  5. How sensitive is TORQUE to FP16 retained values and offline codebook size?
Reviewer_7TtqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TORQUE addresses the loss of quantization accuracy caused by outliers before and after random rotation. It jointly decides which values to retain at higher precision and how to quantize the rest under a shared expected bit budget. The authors show that, for a fixed number of values retained before rotation, choosing the largest-magnitude values minimizes their derived error bound. Experiments report lower reconstruction error across several quantizers and tasks, including activation and KV-cache compression. However, the method adds encoding overhead, and lower reconstruction error does not always improve downstream retrieval accuracy.

Critical strengths
  1. Easy to integrate with existing quantizers. The framework is applied to EDEN, RaBitQ, and TurboQuant, and the paper also analyzes the vector quantizer HIGGS.

  2. Well-motivated problem. The paper identifies that, although random rotation facilitates quantization, it can spread the effect of large values in the original vector across many coordinates.

Critical weaknesses
  1. A gap between the theoretical assumptions and the implementation. The proof relies on direction-independent error under rotation and exact retention, whereas the implementation uses a fast Hadamard rotation and finite-precision values.

  2. NMSE is the main optimization target, rather than downstream performance. The GIST results in the appendix show that Recall@1 can decrease even when reconstruction error improves.

  3. Limited model-level evaluation. The application experiments focus on Qwen2.5-0.5B and do not report generation quality, perplexity, or end-to-end inference performance on larger models.

Questions for authors
  1. Under the randomized Hadamard rotation used in practice, how closely does the optimizer’s predicted NMSE match the measured NMSE? Could any discrepancy change the selected k,c,sk,c,s?

  2. Could the authors report perplexity, generation quality, KV-cache memory usage, and end-to-end throughput on larger language models, in addition to vector reconstruction error?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.