- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper investigates the training of LLMs with binary and ternary weighs without relying on full-precision master weights. The authors claim that prior work on extreme low-bit training still relies on STE with full-precision latent weights that accumulate small continuous updates, so training memory stays close to that of full-precision training. On the other hand, DUET updates the discrete weights directly, converting each Adam step into a per-output-channel transition budget that determines which weights flip, so the learning rate controls how many weights change instead of how much each weight changes. Moreover, they also propose memory-efficient optimizer states, i.e., a block-32 MXINT8 first moment with stochastic rounding, an Adafactor-style factored second moment, and AdamA-style direct gradient accumulation.
Critical strengths
- The paper is clearly written and the mathematical derivatiosn seems correct to me and well-motivated
- DUET keeps the forward pass and the binary/ternary weight format unchanged, so already released QAT checkpoints can be fine-tuned as they are
- The Experimental section is quite extensive. Moreover, the Appendix details most training settings, improving reproducibilty
Critical weaknesses
-
The gains in memory attributed to the removal of master weights are mixed with optimizer-state compression. DUET uses a MXINT8 first moment, a factored second moment, and direct gradient accumulation, while the STE baseline in Table 2 seems to use 8-bit Adam but without the factored second moment and the direct gradient accumulation. Since Table 6 shows that STE works with compressed states, the authors should report STE with the same optimizer-state configuration to isolate the contribution of removing master weights alone.
-
All memory measurements use batch size 1 and sequence length 256, a regime where activations are nearly negligible and weight and optimizer state dominate. This setting maximizes the reported reduction factors, but it is not representative of practical training. The authors should report peak memory at realistic batch sizes and sequence lengths (e.g., the global batch size of 128 used in their fine-tuning experiments and longer contexts) to show how the advantage shrinks as activation memory grows.
-
The fine-tuning evaluation is hard to interpret since Table 1 does not report any row for the CAT-Q ternary checkpoint befor fine-tuning, so the contribution of DUET over the starting point is not clear. Moreover, ternary + LoRA uses a learning rate of , which I think it could be low for LoRA, and the STE full fine-tuning hyperparameters in Table 2 are not reported, so both baselines may be under-tuned.
-
All the reported results are single runs without standard deviations computed on different seeds, so the small gains over baselines could be within evaluation noise. The authors should report experiments with different runs to have more solid results.
-
There in no comparison with other works on direct optimization of discrete weights, even if thre Related Literture section discusses several, e.g., Bop (Helwegen et al., 2019), direct quantized training with stochastic rounding (Zhao et al., 2025) and ECO (Nikdan et al., 2026). The paper should compare against them in the pretraining setting of Table 7.
Questions for authors
- What are the peak memory requirements and accuracy of STE when run using the same MXINT8 first moment, factored second moment and direct gradient accumulation as DUET?
- How do the memory reductions change at realistic batch sizes and sequence lengths?
- Are the gains over STE and Ternary + LoRA statistically significant across seeds and how does DUET compare with Bop or stochastic-rounding direct training at 200M and 1B?