From Sweep to Seam: interleaved cross-block post-training quantization
Source: Kurate cs.AI weekly leaderboard #19, tier 1, absent from HuggingFace · arXiv 2608.09595 · raw
Authors: Achille Jacquemond (Fujitsu), Yuma Ichikawa (Fujitsu / RIKEN Center for AIP), Akira Sakai (Fujitsu / Tokai University).
TL;DR. Post-training quantization of a large model cannot be optimized globally, because the whole model does not fit on one accelerator, so in practice it is done block by block. Everyone has been improving what happens inside the block: better objectives, better weight representations, better error compensation. This paper changes when and in what order the blocks are processed, and reports that the schedule alone is what stands between a usable sub-two-bit model and a broken one. Under a left-to-right sweep of adjacent block pairs, each pair is optimized once and early errors are never revisited, so error compounds with depth. At 1.58-bit ternary quantization that compounding is catastrophic: the paper reports the sequential baseline producing very high perplexity on Llama-3-8B and Qwen3-8B. Interleaved Cross-Block Quantization revisits block interfaces instead of sweeping past them.
flowchart LR
subgraph S[Sequential CBQ baseline]
direction LR
A1[B1-B2] --> A2[B2-B3] --> A3[B3-B4] --> A4[...] --> A5[error<br/>compounds]
end
subgraph I[Interleaved CBQ]
direction LR
C1[B1-B2] --> C2[B2-B3] --> C3[revisit<br/>seam B1-B2] --> C4[B3-B4] --> C5[revisit<br/>seam B2-B3]
end
S --> D{Sub-2-bit<br/>regime}
I --> D
D -->|sequential| E[Perplexity blows up<br/>Llama-3-8B, Qwen3-8B]
D -->|interleaved| F[Usable ternary<br/>1.58-bit model]
classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
class A1,A2,A3,A4,C1,C2,C4 input
class D decision
class C3,C5,F output
class A5,E warn
The mechanism
Quantization here means replacing a model's full-precision weights with a much coarser numeric representation so the model needs less memory and less bandwidth to serve. Post-training quantization does this after training is finished, with no retraining, which is what makes it practical and also what makes it fragile.
The relevant baseline is Cross-Block Quantization, which improved on per-layer methods like GPTQ and AWQ by considering dependencies across several transformer blocks during reconstruction instead of quantizing each in isolation. The paper simplifies CBQ to a fixed two-block weight-only reconstruction objective and calls it Sequential CBQ, then makes its contribution purely at the schedule level.
The problem it names precisely: in a left-to-right sweep over adjacent block pairs, a pair is optimized once. An error introduced in block 2 is then baked into the input distribution that blocks 3 onward are fit against, and nothing ever comes back to correct it. The student's activation stream drifts away from the full-precision teacher's, and the drift accumulates with depth. In moderate-precision regimes this is tolerable. At 1.58-bit ternary it is not, because the per-block error is large enough that compounding dominates.
ICBQ interleaves refinement passes back over the seams between already-quantized block pairs, so critical interfaces are revisited and depth-wise error buildup is actively suppressed rather than merely accounted for. The name is the argument: a sweep passes over each joint once, a seam gets stitched.
How this relates to what the wiki already knows
This is a schedule-level result, and the wiki has been accumulating a case that schedules and orderings are systematically undervalued in efficiency work. MXAttention (08-01) pushed attention itself into MXFP4, and VQ-VLA (08-03) made quantization motion-aware for vision-language-action models. Both change the representation. ICBQ changes nothing about the representation and recovers a model that the same representation otherwise destroys, which is a strong claim about where the remaining headroom is in PTQ. It also means the gain is orthogonal to and composable with the objective-level improvements it cites, QEP's explicit error compensation, LoaQ's output-matching factors, LPCD's relaxed submodule objectives, and with the sub-two-bit weight families OneBit, BitNet, DBF and MDBF.
It is the only Tier 1 efficiency paper on today's board, and it did not come from HuggingFace. The 2026-08-12 HuggingFace batch was agentic and benchmark-heavy with no routing, KV cache, compression or GPU entry. This paper sits at Kurate cs.AI #19 with an ai_rating of 5.0/10, which is mid-pack, and Kurate's tournament had not run this week (every entry at the 1200 baseline with 0% win rate), so the ranking carries little information. That combination, absent from HuggingFace and unranked by Kurate, is exactly the hole this wiki's cross-source rule exists to catch: a directly relevant compression result that neither popularity nor quality signal surfaced.
The interesting connection is to the memory-tiering result from yesterday. OasisKV (08-11) argued that the way to get KV cache capacity is to move the cache to cheaper memory rather than shrink it, because a prefetch miss is a stall while an eviction is a permanent loss. ICBQ is on the other branch, shrinking the weights as aggressively as possible, and it makes the shrink-side branch safer in the same spirit: the failure it fixes is an irreversible one, an error written into the schedule that no later step can undo. Both papers are ultimately about making an efficiency technique's failure mode recoverable rather than permanent. That is a better organizing frame for this page than the shrink-versus-move split alone.
Sub-two-bit matters commercially right now, which is unusual for a quantization schedule paper. The Semiconductor Newsletter's week 32 records OLIX raising $312 million at a $3.3 billion valuation for an HBM-free AI inference architecture, and Sandisk and SK hynix publishing the first open HBF (High Bandwidth Flash) specification for inference memory. An HBM-free or flash-backed inference stack has far less fast memory to work with, which is precisely the regime where ternary weights stop being an academic exercise. Research that makes 1.58-bit reliable and capital that is betting against HBM are pointed at the same deployment.
Gaps
- The paper's own baseline is a simplification it introduced. "Sequential CBQ" is a fixed two-block weight-only reduction of CBQ, not CBQ as published, so the reported blow-up is against a weakened version of the prior method. The comparison against full CBQ is the one that decides how large the contribution is.
- Weight-only. No activation quantization, so this does not yet speak to the weight-activation setting SmoothQuant and its successors target, which is where most serving-throughput gains live.
- Cost of the extra passes is the obvious question and the abstract does not answer it. Interleaved revisiting means more reconstruction compute than a single sweep, and PTQ's whole appeal is being cheap relative to quantization-aware training. If the revisit schedule pushes calibration cost toward QAT territory, the method's positioning changes.
- Two 8B models. No evidence at the scales where block-wise PTQ is not merely convenient but mandatory, which is the setting the paper's own motivation invokes.