When Agents Slow Down: Understanding LLM Agents' Test-Time Strategies via Elo-per-token Analysis

Hugging Face Daily Papers Papers

Summary

This paper introduces Elo-per-token analysis to study how LLM agents allocate test-time compute, revealing that agents initially outperform independent sampling but slow down over time, with parallel sessions offering performance gains.

Large language model (LLM) agents allocate test-time compute adaptively as they revise solutions, use tools, explore alternatives, and decide when to stop. This test-time strategy makes it difficult to measure how agent performance scales. We study open-ended tasks that provide continuous scores for intermediate submissions, making progress observable throughout long trajectories. We propose Elo-per-token analysis, which tracks the best solution found at each token budget and uses a Bradley-Terry model to aggregate within-task orderings into Elo ratings across tasks with different score scales. We apply it to four general-purpose agents on four open-ended benchmarks, with sessions of up to 100M tokens, and to three feedback-driven LLM optimization harnesses in controlled single-task interventions. Independent sampling provides a theoretically characterized reference, for which Elo grows linearly with log compute. Against this reference, agents can initially convert tokens into Elo faster than independent sampling, but their marginal gains diminish and eventually fall below the reference. In contrast, the strongest historical human contestants improve superlinearly over contest time on shared AtCoder Heuristic Contest tasks, providing evidence of continual learning and substantial headroom after agents slow down. We define the scaling inflection point as the per-session budget where marginal Elo gains match the independent-sampling reference. Using this point as the per-session budget, we split 100M tokens across parallel sessions on FrontierCS Polyomino Packing, gaining +264 Elo over one long session and +355 over ten short sessions.
Original Article
View Cached Full Text

Cached at: 09/15/26, 06:40 AM

Paper page - When Agents Slow Down: Understanding LLM Agents’ Test-Time Strategies via Elo-per-token Analysis

Source: https://huggingface.co/papers/2609.15309

Abstract

Elo-per-token analysis reveals that LLM agents initially scale faster than independent sampling but eventually slow, while parallel short sessions improve performance over single long runs.

Large language model (LLM) agents allocatetest-time computeadaptively as they revise solutions, use tools, explore alternatives, and decide when to stop. This test-time strategy makes it difficult to measure how agent performance scales. We studyopen-ended tasksthat provide continuous scores for intermediate submissions, making progress observable throughout long trajectories. We proposeElo-per-token analysis, which tracks the best solution found at each token budget and uses aBradley-Terry modelto aggregate within-task orderings into Elo ratings across tasks with different score scales. We apply it to four general-purpose agents on four open-ended benchmarks, with sessions of up to 100M tokens, and to three feedback-driven LLM optimization harnesses in controlled single-task interventions.Independent samplingprovides a theoretically characterized reference, for which Elo grows linearly with log compute. Against this reference, agents can initially convert tokens into Elo faster thanindependent sampling, but their marginal gains diminish and eventually fall below the reference. In contrast, the strongest historical human contestants improve superlinearly over contest time on sharedAtCoder Heuristic Contesttasks, providing evidence of continual learning and substantial headroom after agents slow down. We define thescaling inflection pointas the per-session budget where marginal Elo gains match the independent-sampling reference. Using this point as the per-session budget, we split 100M tokens across parallel sessions on FrontierCS Polyomino Packing, gaining +264 Elo over one long session and +355 over ten short sessions.

View arXiv pageView PDFAdd to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2609.15309 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.15309 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.15309 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

@polynoamial: https://x.com/polynoamial/status/2064210146558136827

X AI KOLs Following

This article argues that LLM benchmark performance is increasingly a function of test-time compute, and that current evaluation methods fail to capture capability improvements when controlling for inference budget. It advocates for plotting performance vs. tokens, cost, or time, and discusses implications for safety evaluations.

@rohanpaul_ai: New Stanford paper argues that, under equal reasoning budgets, one LLM usually solves multi-hop problems better than ma…

X AI KOLs Timeline

A new Stanford paper shows that under equal reasoning token budgets, single LLMs typically outperform multi-agent systems on multi-hop reasoning tasks, with gains from multi-agent setups often stemming from additional compute rather than architectural superiority. The paper uses the Data Processing Inequality to explain why information loss in handoffs harms multi-agent performance, and identifies context quality as the key factor where multi-agent systems can provide benefits.

LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling

Hugging Face Daily Papers

This paper introduces AutoTTS, an environment-driven framework that automates the discovery of test-time scaling strategies for LLMs by formulating it as controller synthesis. It demonstrates improved accuracy-cost tradeoffs on mathematical reasoning benchmarks with minimal computational overhead.