Evoflux: Inference-Time Evolution of Executable Tool Workflows for Compact Agents

TL;DR — Small language models are cheap to deploy as tool agents, but they generate "plausible" tool-use plans that fall apart in practice: wrong tool resolved, schema violated, a downstream step depends on an output the upstream step never produced. Evoflux reframes compact tool use as workflow repair. At inference time it runs an evolutionary search over typed workflow graphs, mutating them with structured edits and using real execution feedback as the fitness signal. On held-out MCP-Bench tasks (live servers, 250 tools) it raises execution feasibility from roughly 3% to 17-24% across small planners. Crucially, supervised fine-tuning and SFT+DPO on the same search-mined data match, underperform, or collapse below zero-shot, while ReAct hits higher peaks but at much higher variance and token cost.

flowchart LR
  P[Compact planner<br/>draft workflow] --> Pop[Population of<br/>typed graphs]
  Pop --> Edit{Structured<br/>edits}
  Edit --> Run[Execute on<br/>live MCP tools]
  Run -->|feedback| Score[Score:<br/>feasibility]
  Score -->|adaptive intensity<br/>+ diversity prune| Pop
  Score -->|best| Out[Grounded<br/>workflow]
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  class P,Pop input
  class Edit,Score decision
  class Run,Out output

What it is

An inference-time evolutionary search method. It treats an executable tool workflow as a typed graph and evolves a population of candidate graphs through structured edits, execution feedback, adaptive mutation intensity, meta-guided redesign, and diversity pruning. The key claim is that the failure mode of small tool agents — discover tools from a live catalog, satisfy schemas, preserve dependencies across intermediate outputs, ground the final answer in executed evidence — is not fixable by small-corpus distillation. A few hundred teacher traces teach the format but not the recovery behavior needed when a plan fails on a changing tool catalog.

Why it matters

This is the strongest argument yet for execution-grounded search over data-distillation when teacher traces are scarce, and it is squarely in Amit's efficiency interest because the whole point is making compact models reliable. The 3% → 17-24% feasibility jump is large, but the more interesting result is the negative one: SFT+DPO on the same mined data can collapse below zero-shot. That is a direct caution to the standard "mine traces, then fine-tune" pipeline.

Key points

• Compact tool use recast as repair of executable workflow graphs, not one-shot generation.
• Execution feasibility on MCP-Bench rises from ~3% to 17-24% across small planners.
• SFT and SFT+DPO on the same search-mined traces match, underperform, or collapse below zero-shot.
• ReAct reaches higher peaks but with higher variance and token cost; Evoflux is the more reliable option under scarce teacher budgets.

Relation to prior wiki

Same-day companion to HarnessBridge: both target compact-agent reliability, one via the interface, one via the plan. The "evolve at inference time" idea connects to today's EvoArena/EvoMem and EvoBrowseComp under one theme: agents that adapt to changing environments rather than assuming a static one. See the tool-calling and self-evolving-agents concept pages.

Source: HuggingFace Daily Papers · Paper · raw: raw/huggingface/2026-06-14-evoflux-inference-time-evolution-of-executable-tool-workflow.md