Research map / Multi-agent systems and games / LLM agents
Scaling Small Agents Through Strategy Auctions
Read the paperCheck your idea against 108,967 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 SALE, a test-time strategy-auction router where Qwen3 agents of different sizes bid with plans scored by cost and self/peer-jury value, with memory-based refinement, achieving higher pass@1 and lower cost than the largest agent on deep search and coding tasks.
Problem
Small language model agents fail to match large ones as task complexity grows, and existing predictive routers degrade on hard agentic tasks and are costly to train; routing must be cheap and effective on long-horizon tasks.
Approach
Agents bid with short strategic plans in a test-time auction; winners are chosen by minimizing a learned cost-minus-value score (cost from plan length and token price, value from plan entropy and self/peer jury scores), and cheaper agents refine bids using contrastive memory of past auction wins and losses.
Key results
- SALE improves deep search pass@1 to 67.3% vs 63.8% for the best single agent (Qwen3-32B) at lower cost (0.21 vs 0.36 $/Mt on average), with coding 61.1% vs 58.4%.
- SALE reduces reliance on the 32B agent by about 65% on deep search and 40% on coding, and cuts total spend by 42% and 25% respectively.
- On the most complex deep search tasks, the smallest agent reaches only ~25% of the largest agent's pass@1, whereas on the simplest tasks it reaches ~87%.
Surprising findings
- Non-predictive FrugalGPT cascade increased spend and lost accuracy on complex coding tasks.
- Smallest agent's share of workload rose over time as auction memory grew (deep search 3.7% to 11.1%).
Limitations
- Only two domains (deep search and coding) studied
- Only Qwen3 models 4B-32B for the main scaling analysis; no models above 32B
- Memory bank grows linearly and was not studied at large scale or under distribution shift; tool costs not priced
Experimental setup
Datasets: SimpleQA PopQA HotpotQA GAIA Humanity's Last Exam MBPP LeetCode Coding-MCQ
Benchmarks: HST-Bench
Metrics: pass@1 price per million tokens Shapley value
Baselines: Qwen3-32B (best single agent) Willingness-to-Pay router CARROT TensorOpera Router FrugalGPT SALE without memory
Builds on: WTP CARROT FrugalGPT Agent Research Environment
Closest papers by meaning
- Cost-Effective Communication: An Auction-based Method for Language Agent Interaction (AAAI 2026 · similarity 0.89)
- FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · similarity 0.88)
- MonoScale: Scaling Multi-Agent System with Monotonic Improvement (ICML 2026 · similarity 0.87)
- AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.86)
- Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions (NeurIPS 2026 · similarity 0.86)
- More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.86)
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.