AI Paper Atlas

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

Future-Gain Guided Test-Time Learning for Large Language Models

LangYu Bian, Jinwu Hu, Zitian Zhang, Dongjin Yang, Yufeng Wang, Qing Du, Qi Chen, Mingkui Tan

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation for reasoning

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 a future-gain guided test-time learning method for LLMs that addresses error propagation when applying entropy minimization to autoregressive decoding under distribution shift.

Problem

LLMs degrade under distribution shift at deployment; entropy-minimization-based test-time learning on autoregressive generation suffers because early decoding errors propagate to later tokens and updating on them reinforces the drift.

Approach

Guides test-time updates by estimating the future gain of tokens (how they affect downstream generation) rather than applying uniform entropy minimization to all decoded tokens; abstract is truncated so the exact mechanism is not confirmed.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add future-gain token selection (FTS) (ablation)FG-TTL w/o RAA · Six math benchmarks (Llama3.1-8B-Instruct), 4-benchmark subsetaverage accuracy: 39.93 -> 41.87Improved
add risk-aware adaptation (RAA) (ablation)FG-TTL w/o FTS · Six math benchmarks (Llama3.1-8B-Instruct), 4-benchmark subsetaverage accuracy: 39.93 -> 42.01Improved
combine FTS and RAA (ablation)FG-TTL · Six math benchmarks (Llama3.1-8B-Instruct), 4-benchmark subsetaverage accuracy: 41.87 -> 43.56Improved
add RAA on top of FTS (ablation)FG-TTL · Six math benchmarks (Llama3.1-8B-Instruct)average accuracy: 33.82 -> 35.33Improved
add FTS on top of RAA (ablation)FG-TTL · Six math benchmarks (Llama3.1-8B-Instruct)average accuracy: 34.02 -> 35.33Improved
future-gain selection instead of random 20% tokens (ablation)FG-TTL · Six math benchmarks (Llama3.1-8B-Instruct)average accuracy: 33.47 -> 35.33Improved
future-gain selection instead of high-entropy 20% tokens (ablation)FG-TTL · Six math benchmarks (Llama3.1-8B-Instruct)average accuracy: 33.03 -> 35.33Improved
high-entropy token selection instead of random selection (ablation)FG-TTL · Six math benchmarks (Llama3.1-8B-Instruct), 4-benchmark subsetaverage accuracy: 41.29 -> 40.49Hurt
adaptive window instead of best fixed window (N=6) (ablation)FG-TTL · Window size sensitivityaverage accuracy: 34.15 -> 35.33Improved
apply FG-TTL to 4-bit quantized modelLlama3.1-8B-Instruct (4-bit) · Four math benchmarksaverage accuracy: 38.91 -> 40.50Improved
apply TENT/EATA/COME to quantized modelLlama3.1-8B-Instruct (4-bit) · GSM8Kaccuracy: 81.96 -> 64.73Hurt
apply FG-TTL on DomainBenchLlama3.1-8B-Instruct · DomainBenchROUGE-Lsum: 17.31 -> 18.32Improved

Experimental setup

Builds on: entropy minimization test-time learning

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