Independent evaluations will determine whether preserving internal representation geometry during quantization-aware distillation improves NVFP4 model quality over KL-only distillation without reducing low-precision efficiency.

state: expiredheat: lowuncertainty: highconvergesscott: mediumlocal-inference nvfp4 quantization distillation

What is this?

A new arXiv paper proposes CKA-QAD, adding a lightweight layerwise alignment loss to NVFP4 quantization-aware distillation so that a student model preserves the BF16 teacher’s internal representation geometry rather than matching only its output distribution via KL divergence. The authors report that, on Qwen3-4B-Thinking-2507, the method recovers more benchmark accuracy than KL-only QAD—for example, 72.3% versus 68.5% on AIME25—after NVFP4 quantization. The supplied results do not identify the authors or provide independent replication, and they do not directly establish that the added training objective preserves the same deployed inference efficiency, beyond both approaches producing NVFP4 models.

Why it matters to Scott

CKA-QAD gives a concrete model-compression analogue of Scott’s Derivational Provenance position: matching the final answer does not establish that the underlying route or representation remains intact. It also bears directly on his hardware-aware local-inference work by proposing a potentially better quality-preservation objective for NVFP4 deployment, but independent evaluation and unchanged runtime efficiency remain unestablished.

ip:framework.derivational-provenanceip:concept.evaluation-driven-developmentdev:concept.hardware-aware-local-inferenceradar:compressed-llm-fidelity-safety-gapradar:concept.quantizationradar:concept.model-compressionradar:concept.model-distillationradar:concept.model-evaluationradar:concept.inference-efficiency

queries asked of Scott's wikis

quantization quality beyond output accuracy
internal representations as model evaluation signals
local inference precision-quality tradeoffs
distillation objectives for compressed models
independent evaluation of quantized models
NVFP4 hardware and inference strategy

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	[2606.05682] Beyond Output Matching: Preserving Internal Geometry in NVFP4 LLM Distillation LocalLLaMA	Aaaaaaaaaeeeee	24	5
🟧 echo.paper ⭐	The authors’ original arXiv paper introduces CKA-QAD, arguing that “output matching alone can mask internal degradation” and proposing CKA r	Fangbo Tu, Junhua Zhao, Chi Liu, Xin Chen, Haifeng Wu, Jian Wan, Srinivasan Manoharan	—	—

Interpretation history

2026-08-14T00:35:50Z
No independent reproduction, implementation, or efficiency comparison emerged across repeated staleness checks, so this remains an isolated author-reported result with no active momentum.
2026-08-12T00:22:50Z
The staleness check found no independent evaluation, implementation, or efficiency measurement; the case remains an unreplicated author-reported proposal rather than an emerging result.
2026-08-09T23:28:47Z
The refreshed discussion raises a prior-work and method-complexity objection, but supplies neither an independent evaluation nor an implementation comparing CKA-QAD with KL-only or layerwise distillation. The case therefore remains an unreplicated author-reported proposal.
2026-08-09T21:33:17Z
The forced re-evaluation adds no substantive evidence: CKA-QAD remains an unreplicated author-reported result, with neither independent quality comparisons nor confirmation of unchanged inference efficiency.
2026-08-09T21:26:32Z
grounded: converges/medium — CKA-QAD gives a concrete model-compression analogue of Scott’s Derivational Provenance position: matching the final answer does not establish that the underlyin
2026-08-09T21:23:47Z
origin walked (codex/luna, conf 0.99): anchor reddit.post.1vk08zl -> echo.paper.8b3a380fb0 by Fangbo Tu, Junhua Zhao, Chi Liu, Xin Chen, Haifeng Wu, Jian Wan, Srinivasan Manoharan
2026-08-09T21:23:16Z
case created — The linked paper introduces a concrete low-precision distillation method, but it has only one lightly engaged observation and no independent validation.

Decision trace

08-14 10:35expireNo independent reproduction, implementation, or efficiency comparison emerged across repeated staleness checks, so this remains an isolated author-reported result with no active momentum.
08-14 10:35alert_silentThe latest trigger is only elapsed staleness and adds no consequential evidence; revive the case only if an independent quality and deployment-efficiency comparison appears.
08-14 10:35alert_routeThe latest trigger is only elapsed staleness and adds no consequential evidence; revive the case only if an independent quality and deployment-efficiency comparison appears.
08-12 10:22repriceThe staleness check found no independent evaluation, implementation, or efficiency measurement; the case remains an unreplicated author-reported proposal rather than an emerging result.
08-12 10:22alert_silentNo consequential evidence has arrived since the prior review; wait for an independent reproduction or measured quality, runtime, and memory comparison.
08-12 10:22alert_routeNo consequential evidence has arrived since the prior review; wait for an independent reproduction or measured quality, runtime, and memory comparison.
08-10 20:21sensor_dirtyengagement_update
08-10 12:21sensor_dirtyengagement_update
08-10 09:28repriceThe refreshed discussion raises a prior-work and method-complexity objection, but supplies neither an independent evaluation nor an implementation comparing CKA-QAD with KL-only or layerwise distillat
08-10 09:28alert_silentA Reddit comment questioning novelty and suggesting layerwise progressive distillation does not materially establish or refute the paper’s quality and efficiency claims; this can wait for measured com
08-10 09:28alert_routeA Reddit comment questioning novelty and suggesting layerwise progressive distillation does not materially establish or refute the paper’s quality and efficiency claims; this can wait for measured com
08-10 09:21sensor_dirtycomment_update
08-10 07:33repriceThe forced re-evaluation adds no substantive evidence: CKA-QAD remains an unreplicated author-reported result, with neither independent quality comparisons nor confirmation of unchanged inference effi
08-10 07:33alert_silentOnly an unchanged re-observation occurred; there is no new consequential delta to surface before a normal briefing. Wait for an independent reproduction, implementation, or measured runtime and memory
08-10 07:33alert_routeOnly an unchanged re-observation occurred; there is no new consequential delta to surface before a normal briefing. Wait for an independent reproduction, implementation, or measured runtime and memory
08-10 07:30alert_silentThe original paper is a substantive, Scott-relevant proposal, but the only new evidence is the authors’ own claim that CKA-guided alignment improves NVFP4 distillation. Independent quality comparisons
08-10 07:30surface_candidateThe original paper is a substantive, Scott-relevant proposal, but the only new evidence is the authors’ own claim that CKA-guided alignment improves NVFP4 distillation. Independent quality comparisons
08-10 07:30alert_routeThe original paper is a substantive, Scott-relevant proposal, but the only new evidence is the authors’ own claim that CKA-guided alignment improves NVFP4 distillation. Independent quality comparisons
08-10 07:26groundCKA-QAD gives a concrete model-compression analogue of Scott’s Derivational Provenance position: matching the final answer does not establish that the underlying route or representation remains intact
08-10 07:23promote_anchororigin walk conf 0.99
08-10 07:23createThe linked paper introduces a concrete low-precision distillation method, but it has only one lightly engaged observation and no independent validation.