Skip to the content.

Recurrent Residual Quantization evaluation

Named work: Recurrent Residual Quantization: A Progressive Multi-Precision Representation for LLMs
Disposition: ABSTAIN from integration until a checkpoint and executable runtime are independently pinned. The metadata contract is suitable to MONITOR and can later DELEGATE execution to a pinned external runtime; fak does not claim an RRQ kernel.

Pinned source and reproduction

Item Pin
Paper arXiv:2608.04048v1https://arxiv.org/abs/2608.04048v1
Public PDF https://arxiv.org/pdf/2608.04048v1
PDF SHA-256 08d39e39738289bcd2784364a713ed337dc4bca6467df8e81046dfb2cf6dcd45
Recipe rrq-paper-algorithm-1/v1: 2-bit PTQ-or-RTN base, successive 2-bit RTN residuals, effective tiers [2,4,6,8], calibration-free
Artifact Paper only. No released weight URI or weight digest was available in the reviewed v1 materials.
Runtime external-rrq-runtime is a delegation boundary, not an implementation. No public package/version/ABI was available in v1; the abstract says code will be released upon publication.
Review date 2026-08-10

Bounded public reproduction (PowerShell):

Invoke-WebRequest https://arxiv.org/pdf/2608.04048v1 -OutFile 2608.04048v1.pdf
(Get-FileHash 2608.04048v1.pdf -Algorithm SHA256).Hash.ToLower()
go test ./internal/residualquant -run 'TestNamedResearchWitness|TestResearchEvaluationSeparatesEvidence|TestResearchDocumentPinsContract'

The expected hash is the pin above. The Go witness independently reads this document and exercises at least three outcomes, including unknown contract, invalid tier, absent executable artifact, and a fully pinned delegate fixture.

What the paper actually evaluates

RRQ stores a low-bit base plus recurrent quantized residual corrections. A prefix reconstructs an effective precision: base only is 2-bit, then one, two, or three 2-bit residuals produce 4-, 6-, or 8-bit weights. The authors evaluate six model families, use WikiText-2 perplexity and eight LM Evaluation Harness zero-shot tasks, and report quantization experiments on 8 NVIDIA A100 80GB GPUs. Algorithm 1 is the recipe pinned above.

The following boundaries are mandatory:

No lab GPU run is warranted for the present acceptance gate: there is no independently identifiable RRQ artifact or runtime to execute. Hardware absence is not the blocker; missing public executable inputs are. Once those inputs exist, the run must use the sanctioned lab/private-comms route and record artifact digest, recipe ID, runtime/version, device, commands, raw outputs, and observed quality/performance separately.

Typed capability contract

internal/residualquant exposes fak.residualquant/v1 and never silently chooses another precision or kernel.

Request Outcome Reason
Inspect the exact paper/recipe pin and a tier in [2,4,6,8] supported RRQ_METADATA_OK
Unknown contract/method, malformed tier chain, unpinned paper/recipe, or unavailable operation unsupported Specific RESIDUALQUANT_* reason code
Execute without a weight URI + digest unsupported RESIDUALQUANT_ARTIFACT_UNPINNED
Execute without runtime ID + version + device unsupported RESIDUALQUANT_RUNTIME_UNPINNED
Execute with all artifact, recipe, and runtime pins delegate RRQ_RUNTIME_HANDOFF and the exact runtime ID

The delegate result says only that the request is sufficiently identified to hand to an external runtime. It does not say the runtime supports the tensors, that quality is preserved, or that performance improves.

Artifact tiers and cache transport decision

Integrate metadata, abstain on executable transport. A cache key may model a tier as:

(method=rrq, recipe=rrq-paper-algorithm-1/v1,
 base-artifact-digest, ordered-residual-digests, prefix-count,
 runtime-id, runtime-version, device)

The ordered residual digest list and prefix count are semantic: omitting either could alias 2-, 4-, 6-, and 8-bit reconstructions. The paper’s additive representation makes this mapping plausible, but v1 does not define a stable interchange container or runtime ABI. Consequently fak must not create a fak-only weight format or treat cache transport as executable support.

Promotion gate

Change the disposition from ABSTAIN to integration only when all are independently readable:

  1. an official artifact URI and cryptographic digest;
  2. an exact conversion recipe/version and ordered residual layout;
  3. a runtime/package version and supported device envelope;
  4. a clean-room command that loads every claimed tier without hidden fallback;
  5. observed quality results against named baselines and datasets; and
  6. observed construction/load/serve measurements from sanctioned compute, with raw evidence.

Until then, monitor the source release and return typed unsupported/delegate outcomes rather than inferring support.