AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

ARC-Decode: Accelerated Decoding with Risk-Bounded Acceptance

Ying Li, Zhaode Wang, Zhiwen Chen, chengfei lv, Huan Wang

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Speculative decoding

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ARC-Decode, a speculative decoding variant that uses risk-bounded acceptance to relax the over-rejection of the classical lossless verification rule under sampling.

Problem

Speculative decoding speedups under sampling (T>0) lag behind greedy decoding because the classical lossless verification rule over-rejects draft tokens.

Approach

Replaces the classical lossless acceptance rule with a risk-bounded acceptance criterion that accepts more draft tokens under sampling while bounding the resulting risk; details beyond the abstract are not stated.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add entropy-guided pruning to LTS-only ARC (ablation)ARC-Decode (Llama-3.1-8B) · MT-Benchspeedup: 2.28× -> 2.61×Improved
add entropy-guided pruning to LTS-only ARC (ablation)ARC-Decode (Llama-3.1-8B) · Alpacaspeedup: 2.06× -> 2.28×Improved
add entropy-guided pruning to LTS-only ARC (ablation)ARC-Decode (Llama-3.1-8B) · HumanEvalspeedup: 2.08× -> 2.13×Mixed
add LTS-only relaxed acceptance to EAGLE-3 (ablation)EAGLE-3 (Llama-3.1-8B) · Alpacaspeedup: 1.42× -> 2.06×Improved
add LTS-only relaxed acceptance to EAGLE-3 (ablation)EAGLE-3 (Llama-3.1-8B) · MT-Benchspeedup: 1.84× -> 2.28×Improved
use entropy-guided pruning instead of depth-based pruning (ablation)ARC-Decode (Qwen3-8B) · MT-Benchaccept length: 2.53 -> 2.70Improved
use entropy-guided pruning instead of depth-based pruning (ablation)ARC-Decode (Qwen3-8B) · MT-Benchtokens/s: 51.2 -> 56.4Improved
use entropy-guided pruning instead of cumulative-probability pruning (ablation)ARC-Decode (Qwen3-8B) · MT-Benchtokens/s: 51.8 -> 56.4Improved
raise sampling temperature from 0.1 to 0.9 (ablation)ARC-Decode (Llama-3.1-8B) · MT-Benchspeedup: 3.13× -> 2.66×Hurt
raise LTS threshold theta from 0.3 to 0.9 (ablation)ARC-Decode (Llama-3.1-8B) · four tasksspeedupHurt
replace EAGLE-3 with ARC-DecodeLlama-3.1-8B · HumanEvalpass@10 (%): 69.9 -> 68.2No clear effect
replace EAGLE-3 with ARC-DecodeLlama-3.1-8B · MT-BenchGPT-4o score: 6.77 -> 7.02No clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Builds on: speculative decoding

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy