---
type: "post"
id: "built-on-gsq-and-rco-methods-from-ista-c27305"
slug: "built-on-gsq-and-rco-methods-from-ista-c27305"
title: "Built on GSQ and RCO methods from ISTA-DASLab with Swift-specific refinement"
description: "Thesis: Learned scalar quantization may recover much of the accuracy gap to more complex vector methods while retaining compatibility with scalar inference kernels."
retrieval_nugget: "Thesis: Learned scalar quantization may recover much of the accuracy gap to more complex vector methods while retaining compatibility with scalar inference kernels."
published_at: "2026-10-12"
updated_at: "2026-10-12"
record_date: "2026-10-12"
date_kind: "published_at"
topics: []
source_urls: ["https://arxiv.org/abs/2604.18556"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:built-on-gsq-and-rco-methods-from-ista-c27305"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:aa75d1ace9aab3d5a88bf82c40ac329d8f839f4e082801d7f916001daeca99ab","reviewed_at":"2026-10-12T17:02:21.179Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Thesis: Learned scalar quantization may recover much of the accuracy gap to more complex vector methods while retaining compatibility with scalar inference kernels.","observed_facts":[{"text":"The paper says GSQ jointly learns per-coordinate grid assignments and per-group scales using a Gumbel-Softmax relaxation.","source_urls":["https://arxiv.org/abs/2604.18556"]},{"text":"The authors report closing most of the gap to QTIP at 2 and 3 bits on Llama 3.1 8B and 70B Instruct.","source_urls":["https://arxiv.org/abs/2604.18556"]},{"text":"The paper describes applying GSQ to public GGUF checkpoints and projecting outputs back into that deployment format.","source_urls":["https://arxiv.org/abs/2604.18556"]}],"mechanism":"Mechanism: GSQ uses a Gumbel-Softmax relaxation to optimize discrete scalar grid assignments alongside group scales, then uses a symmetric scalar grid with group-wise quantization.","why_now":"Why now: Low-bit deployment is common for local inference, yet scalar methods are described as accuracy-limited and more complex vector or trellis methods as difficult to implement and scale.","implications":["First implication: Deployment teams can test a potential low-bit quality improvement without changing the scalar inference-kernel interface.","The reported accuracy gains need to be weighed against measured runtime and memory on target workloads."],"evidence_boundary":"Evidence boundary: The supplied arXiv paper reports the experiments and claims; it does not independently establish workload-level speed, memory savings, or generalization across models.","watch_conditions":["First watch condition: Whether independent evaluations reproduce the quality gains while measuring runtime and memory on real deployment workloads."],"related_records":[],"new_branch_reason":"No prior New Runtime record is supplied; this paper supports a distinct research record about quantization methods and low-bit deployment."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/built-on-gsq-and-rco-methods-from-ista-c27305/","markdown":"https://newruntime.com/posts/built-on-gsq-and-rco-methods-from-ista-c27305.md","json":"https://newruntime.com/posts/built-on-gsq-and-rco-methods-from-ista-c27305.json"}
---

# Built on GSQ and RCO methods from ISTA-DASLab with Swift-specific refinement

## Retrieval answer

Thesis: Learned scalar quantization may recover much of the accuracy gap to more complex vector methods while retaining compatibility with scalar inference kernels.

Thesis: Learned scalar quantization may recover much of the accuracy gap to more complex vector methods while retaining compatibility with scalar inference kernels. The arXiv paper reports that GSQ jointly learns per-coordinate grid assignments and per-group scales using a Gumbel-Softmax relaxation. It says the method closes most of the gap to QTIP at 2 and 3 bits on Llama 3.1 8B and 70B Instruct.

Mechanism: GSQ uses a Gumbel-Softmax relaxation to optimize discrete scalar grid assignments alongside group scales, then uses a symmetric scalar grid with group-wise quantization. The authors also describe applying the approach to public GGUF checkpoints and returning the result to that format. Why now: Low-bit deployment is common for local inference, yet scalar methods are described as accuracy-limited and more complex vector or trellis methods as difficult to implement and scale. The contribution targets that tension directly.

First implication: Deployment teams can test a potential low-bit quality improvement without changing the scalar inference-kernel interface. First watch condition: Whether independent evaluations reproduce the quality gains while measuring runtime and memory on real deployment workloads. Evidence boundary: The supplied arXiv paper reports the experiments and claims; it does not independently establish workload-level speed, memory savings, or generalization across models. The paper further reports scaling to Kimi-K2.5; that result is not, by itself, proof of production readiness.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)