acceptodds
Under review as a conference paper at ICLR 2027

Ternova: Adaptive Low-Rank Post-Training Ternarization for Large Language Models

Abstract

Large language models (LLMs) have achieved remarkable capabilities, yet their massive memory footprint and inference cost hinder efficient deployment. Ternary post-training quantization (PTQ) offers an attractive solution by combining compact storage with richer representation capacity than binary quantization. However, existing ternary PTQ methods remain challenged by limited ternary representation capacity, heterogeneous structural sensitivity across Transformer modules, and the mismatch between local reconstruction and model-level objectives. To address these challenges, we propose Ternova, an adaptive low-rank post-training ternarization framework for LLMs. Ternova is built on three core components: 1) Hessian-Aware Annealed Ternary Factorization (HATF) constructs high-fidelity low-rank ternary representations through sensitivity-aware factorization and annealed discrete refinement; 2) Sensitivity-Reweighted Mixed-Rank Allocation (SRMA) adaptively distributes rank capacity according to structural irreplaceability, task sensitivity, and storage cost; and 3) Task-Aligned Ternary Refinement (TATR) further aligns the compressed model with task-level behavior through global scale calibration and sparse topology refinement. Ternova preserves a compact ternary deployment form without additional dense high-precision branches. Extensive experiments across diverse LLM families demonstrate that Ternova achieves competitive or superior performance over existing low-bit PTQ methods at only 1.59 BPW while retaining competitive reasoning capabilities. On Qwen3.5-27B, Ternova achieves 74.62% average zero-shot accuracy, approaching the 75.97% BF16 baseline. Moreover, Ternova achieves roughly 3 inference speedup and up to 80.6% memory savings over BF16, demonstrating a strong balance among accuracy, compression, and practical efficiency.

Primary area:
foundation or frontier models, including LLMs
Keywords:
post-training quantization, ternarization, LLMs

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_vn7SOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Ternova combines Hessian-aware ternary factorization, sensitivity-based mixed-rank allocation, and task-aligned refinement to compress pretrained LLMs without a dense high-precision residual branch. It evaluates model quality and efficiency across several families, reporting approximately 1.59 bits per quantized linear weight.

Critical strengths

The three stages address plausible sources of compression error within a coherent framework. The evaluation covers several model families, and Table 6 provides a useful storage-matched DBF+PV comparison with improvements on various models.

Critical weaknesses
  1. DBF (https://arxiv.org/html/2505.11076v4) already uses scaled discrete factors, nonuniform intermediate dimensions, and global refinement. Appendix A.4.4 explains Ternova's differences, Table 6 matches storage, and Tables 3, 4, and 7 provide component evidence. However, matched optimization costs for DBF+PV are not established, and Table 3 does not report each variant's realized storage and compute budget despite the use of a higher-capacity probe. These omissions limit attribution of the improvements to the proposed mechanisms rather than additional capacity or optimization.
  2. Section 4.4 names the GPU and llama.cpp but does not specify the inference batch size, prompt/generation lengths, prefill/decode distinction, or implementation of the two ternary factors.
  3. The QAT timing comparison also uses different storage budgets: approximately 1.03–1.04 BPW versus Ternova's 1.59. Table 2 lacks sufficient generation/scoring settings, while HumanEval drops from 78.05 to 49.36, qualifying the broad capability-retention claim.
Questions for authors
  1. Under matched data, realized storage, and optimization cost, what gains remain over DBF+PV and a direct ternary extension? Please specify the budgets for the existing component ablations.
  2. What exact implementation and workloads produced Figure 3, and were the timing comparisons measured on comparable hardware with all preprocessing included?
  3. Can you provide Table 2's generation/scoring settings, per-model packed storage, and a correction or explanation for Table 5's average?
Reviewer_uNiNOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This methodological study focuses on post-training ternarization (PTQ) for LLMs. To address the limited representational capacity of a single binary/ternary discrete matrix in existing PTQ methods (which often rely on auxiliary parameters like fine-grained scales that increase actual storage overhead), heterogeneous sensitivity across Transformer modules, and the mismatch between local reconstruction and global task objectives, the authors propose the Ternova framework. It approximates linear layer weights as the product of two low-rank ternary matrices and three 1D scaling vectors: W^=Diag(a)TADiag(m)TBDiag(b)\hat{W} = \mathrm{Diag}(a) T_A \mathrm{Diag}(m) T_B \mathrm{Diag}(b), which aims to increase representation capacity while keeping scaling parameter storage overhead low.

The framework consists of three modules:

  1. HATF construct a curvature-aware space W′=DoWDiW' = D_o W D_i using diagonal K-FAC statistics, obtain initial ternary factor via TSVID projection and ADMM. Then jointly refines TA,TBT_A, T_B and scaling vectors via mask-and-sign annealing.
  2. SRMA evaluates refit-aware deletion loss in the Gram space and greedily allocates layer rank capacity under a global BPW budget by combining task sensitivity and storage cost.
  3. TATR applies end-to-end token-level knowledge distillation to globally calibrate scaling vectors and sparsely update the discrete factors via STE under adjacent-state transition constraints, mitigating cumulative errors from layer-wise factorization.

Experiments are conducted on LLaMA-3, Qwen3/3.5/3.8, and Gemma-3/4 models. In Table 1, the quantized linear layers have an effective bit-width of 1.59 BPW. Ternova outperforms low-bit PTQ baselines at 2.28–3.50 BPW on 7 tasks zero-shot average accuracy for most models and ranks in the top two on WikiText2 PPL. It also shows clear advantages on complex reasoning and coding tasks including MMLU, GSM8K, and HumanEval. Regarding efficiency, quantizing LLaMA-3-8B takes 9.2 GPU hours, saving over 30x in time cost compared to QAT methods (e.g., LittleBit, OneBit) while achieving better PPL.

Critical strengths

S1. The research problem is of practical importance. It focuses on the trade-off among representational capacity, auxiliary parameter storage overhead, and target task performance in ultra-low-bit quantization, with a clear motivation. S2. The method is specifically designed for the problem, with clear and rigorous mathematical derivations. In HATF, the diagonal second-order approximation greatly reduces calculation cost and is encoded into W′W' beforehand via reparameterization, maintaining zero additional deployment overhead. Evaluating refit-aware deletion loss in the low-dimensional Gram space in SRMA, as well as sparse updates constrained to adjacent state transitions in TATR, effectively balances computational efficiency and optimization stability. S3. The experiments provide relatively complete evidence for the overall effectiveness of the framework. The method achieves good accuracy and storage compression across multiple models and tasks, outperforming DBF+PV under the same 1.59 BPW budget. Component ablations further support the contribution of each module under the tested settings.

Critical weaknesses

W1. The experimental baselines lack recent ternary PTQ methods. Section 2.2 mentions three representative recent ternary PTQ works (PT2-LLM\text{PT}^2\text{-LLM}, PTQTP, and TWLA). However, in Section 4, Table 1 only includes TWLA as a ternary baseline and lacks results for the other two methods. Furthermore, Table 2 (evaluating complex reasoning on MMLU, HumanEval, and GSM8K) does not include any ternary PTQ baselines (not even TWLA), which is insufficient to establish Ternova's advantage over existing ternary methods on these tasks. W2. There is a lack of controlled experiments to identify the true source of gains from the ternary representation over the binary representation. The K-FAC importance weighting and ADMM dual-factor decomposition in HATF are heavily inherited from the binary factorization method DBF. Under the same BPW budget, binary factorization naturally allows a larger intermediate rank dimension. Thus, although ternarization introduces an extra discrete state, it does not automatically guarantee stronger representational capacity at the same storage budget. While Appendix A.4.4 provides an equal-budget comparison with DBF+PV, it changes both the discrete representation and the optimization strategy simultaneously; current ablations do not adequately separate the gains of the ternary representation itself from those of the newly added optimization components.

Questions for authors

Q1. Regarding W1, please provide the effective BPW, WikiText2 PPL, and 7-task zero-shot average accuracy for PT2-LLM\text{PT}^2\text{-LLM} and PTQTP under comparable settings, and include at least one ternary PTQ baseline (such as TWLA) in Table 2. If these results cannot be provided, please justify why the current comparisons are sufficient to support your claims or clarify the scope of your conclusions. Q2. Regarding W2, could you provide a controlled comparison between binary and ternary factorization under the exact same BPW budget and optimization pipeline (for instance, comparing ADMM-TSVID directly against ADMM-SVID under the shared optimization workflow) to separate the contribution of changing the discrete representation (ternary vs. binary) from that of the newly introduced optimization strategies?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.