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

• URL: http://arxiv.org/abs/2609.15309v1
• Date: Mon, 14 Sep 2026 09:57:49 GMT
• Title: When Agents Slow Down: Understanding LLM Agents' Test-Time Strategies via Elo-per-token Analysis
• Abstract summary: Large language model (LLM) agents allocate test-time compute adaptively as they revise solutions, use tools, explore alternatives, and decide when to stop.<n>We propose Elo-per-token analysis, which tracks the best solution found at each token budget.<n>We apply it to four general-purpose agents on four open-ended benchmarks, with sessions of up to 100M tokens.
• Score: 66.34225048563418
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• PACE: A Proxy for Agentic Capability Evaluation [60.3743414796937]
  We introduce PACE, a framework that constructs proxy benchmarks by selecting instances from existing non-agentic evaluations.<n>Experiments across 14 models, 4 agentic benchmarks, and 19 non-agentic benchmarks show that PACE-Bench predicts agentic scores with leave-one-out cross-validation (LOOCV) mean absolute error (MAE) under 4%, Spearman correlation above 0.80, and pairwise model-ranking accuracy around 85%, all at much less than 1% of the full agentic evaluation cost.
  arXiv  Detail & Related papers  (2026-07-02T10:59:03Z)
• IdleSpec: Exploiting Idle Time via Speculative Planning for LLM Agents [56.77101184011525]
  Large language model (LLM)-based agents solve complex tasks by leveraging multi-step reasoning with iterative tool calls and environment interactions.<n>Despite the prevalence of idle time in most agentic scenarios, existing works treat it as an unavoidable overhead.<n>We introduce IdleSpec, a scalable and generic inference approach that leverages idle-time computation to improve agent performance.
  arXiv  Detail & Related papers  (2026-05-21T08:25:17Z)
• AEL: Agent Evolving Learning for Open-Ended Environments [43.56685432981852]
  We introduce emphAgent Evolving Learning (ael), a two-timescale framework that addresses this obstacle.<n>ael achieves a Sharpe ratio of 2.13$pm$0.47, outperforming five published self-improving methods.<n>This demonstrates that the bottleneck in agent self-improvement is emphself-diagnosing how to use experience rather than adding architectural complexity.
  arXiv  Detail & Related papers  (2026-04-23T14:29:25Z)
• Agentic Aggregation for Parallel Scaling of Long-Horizon Agentic Tasks [33.36477179106684]
  We study parallel test-time scaling for long-horizon agentic tasks such as agentic search and deep research.<n>We propose AggAgent, an aggregation agent that treats parallel trajectories as an environment.
  arXiv  Detail & Related papers  (2026-04-13T17:26:31Z)
• What If We Allocate Test-Time Compute Adaptively? [2.1713977971908944]
  Test-time scaling allocates inference computation uniformly, uses fixed sampling strategies, and applies verification only for reranking.<n>We propose a verifier-guided adaptive framework treating reasoning as iterative trajectory generation and selection.<n>Across datasets, our dynamic, PRM-guided approach consistently outperforms direct test-time scaling.
  arXiv  Detail & Related papers  (2026-02-01T07:30:22Z)
• A BERTology View of LLM Orchestrations: Token- and Layer-Selective Probes for Efficient Single-Pass Classification [2.0069888187253615]
  Production LLM systems often rely on separate models for safety and other classification-heavy steps.<n>We instead reuse computation already paid for by the serving LLM: we train lightweight probes on its hidden states and predict labels in the same forward pass used for generation.
  arXiv  Detail & Related papers  (2026-01-19T18:40:29Z)
• AgentPRM: Process Reward Models for LLM Agents via Step-Wise Promise and Progress [71.02263260394261]
  Large language models (LLMs) still encounter challenges in multi-turn decision-making tasks.<n>We build process reward models (PRMs) to evaluate each decision and guide the agent's decision-making process.<n>AgentPRM captures both the interdependence between sequential decisions and their contribution to the final goal.
  arXiv  Detail & Related papers  (2025-11-11T14:57:54Z)
• MM-HELIX: Boosting Multimodal Long-Chain Reflective Reasoning with Holistic Platform and Adaptive Hybrid Policy Optimization [103.74675519953898]
  Long-chain reflective reasoning is a prerequisite for solving complex real-world problems.<n>We build a benchmark consisting 1,260 samples of 42 challenging synthetic tasks.<n>We generate post-training data and explore learning paradigms for exploiting such data.
  arXiv  Detail & Related papers  (2025-10-09T17:53:58Z)
• Multi-Agent Sampling: Scaling Inference Compute for Data Synthesis with Tree Search-Based Agentic Collaboration [81.45763823762682]
  This work aims to bridge the gap by investigating the problem of data synthesis through multi-agent sampling.<n>We introduce Tree Search-based Orchestrated Agents(TOA), where the workflow evolves iteratively during the sequential sampling process.<n>Our experiments on alignment, machine translation, and mathematical reasoning demonstrate that multi-agent sampling significantly outperforms single-agent sampling as inference compute scales.
  arXiv  Detail & Related papers  (2024-12-22T15:16:44Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.