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

The gist The paper introduces Evoflux, an inference-time evolutionary search method that treats compact tool use as the repair of executable tool workflows, showing that execution-grounded search is more reliable under scarce teacher-trace budgets than supervised finetuning or SFT+DPO on the same data.

Problem Setup

Small language models often generate plausible workflow graphs that fail under tool resolution, parameter validation, dependency tracking, or execution The Model Context Protocol (MCP) offers a concrete interface for this class of agents by standardizing how models discover and call tools exposed by external servers Compact language models in the 1.5B to 4B range are attractive for such agents because they reduce serving cost, latency, and privacy exposure while supporting local or enterprise-controlled deployment They emit malformed JSON, select unavailable tools, omit required arguments, break dependency links, and answer from prior knowledge rather than executed evidence

Method Overview

Evoflux is an adaptive evolutionary loop that searches for executable tool workflows at inference time Given task q = (x, T, E), a planner P proposes a symbolic workflow The candidate c stores everything needed to evaluate one workflow variant The procedure BUILDCANDIDATE converts a symbolic graph into an executable workflow Scoring function f(c) maps each candidate to a scalar execution score derived from the MCP-Bench LLMJUDGE evaluator

Key Components of Evoflux

Evoflux evolves plans through several mechanisms The process involves structured edits, execution feedback, adaptive intensity control, meta-guided redesign, and action-hash diversity pruning

** Typed edit language modifies workflows through operations like Tool swaps, Parameter edits, Tool insertion, Tool removal, and Step reordering Mutation uses this evidence to choose edit locations and edit types The controller tracks recent positive progress through a growth statistic G; the intensity I increases when progress stalls and decreases when improvements accumulate Meta guidance supplies a higher-level redesign step when local mutation stops improving the population The population is capped at K candidates, and survivors are chosen using action-hash binning to preserve high-scoring candidates while reducing collapse onto near-duplicate workflows The heuristic fallback method provides a low cost safety path when the planner cannot call or successfully use the language model This fallback uses token overlap between the query and the registered tools to select at most min(self.max steps, 3) tools in a strictly sequential graph The method returns a GraphTemplate marked with heuristic=True This fallback acts as a deterministic execution floor It can keep the system moving when credentials are missing or model calls fail, but it cannot reason over schemas, task semantics, or multi-hop evidence with the same flexibility as the model-based planner The full procedure is detailed in Algorithm 2 This algorithm maintains a bounded population P, the best observed score f∗, and an adaptive growth statistic G At each step, the current growth level determines the search intensity Low growth increases exploratory mutation and may trigger meta guidance High growth concentrates search around stronger candidates through tournament-style parent selection and evidence-conditioned repair The procedure BUILDCANDIDATE converts a symbolic graph into an executable workflow It parses the graph, resolves server and tool identifiers against the catalog T, validates parameters against tool schemas, checks graph acyclicity and dependency references, and executes the workflow when static checks succeed The scoring function f(c) maps each candidate to a scalar execution score derived from the MCP-Bench LLMJUDGE evaluator Infeasible candidates receive a score of 0 For feasible candidates, the evaluator is run with stability testing enabled, invoking it five times per candidate with a freshly randomized prompt structure on each pass The scalar score drives parent selection, growth updates, population pruning, and final candidate selection The search split shows how often each planner reaches executable workflows, how much the evolutionary loop improves over the initial proposal, where the judge’s sub-dimensions change after search, and how score trades off against token cost Evoflux raises execution feasibility from roughly 3% to 17–24% across small planners In contrast, SFT and SFT+DPO on the same search-mined data match, underperform, or collapse below zero-shot performance This results show that executiongrounded search is more reliable under scarce teacher-trace budgets The validation split shows that Valid Best recovers roughly the same score range as Best for Llama-3.2-3B and roughly two-thirds of the Best score for Qwen3.

1. Bold header: Inference-time workflow repair for compact agents

This system can treat compact tool use as the repair of executable tool workflows by evolving typed workflow graphs through structured edits, execution feedback, adaptive intensity control, meta-guided redesign, and diversity pruning. This allows the model to recover relevant tools from a changing catalog and satisfy schemas during live inference.

2. Bold header: Quantitative comparison of training versus search

The system can determine if it is better to spend compute on weight updates, or on inference-time repair over the actual task, catalog, and execution environment. This addresses the risk that small-corpus finetuning may carry downside risk for compact models by showing that inference-time search lifts held-out feasibility and score across small planners.

3. Bold header: Robustness against scarce teacher trace budgets

The system is designed to perform reliably with limited supervision, as evidenced by results showing that SFT and SFT+DPO on the same search-mined data match, underperform, or collapse below zero-shot performance. This suggests that executiongrounded search is more reliable under scarce teacher-trace budgets than relying solely on supervised finetuning.

4. Bold header: Improved structural generalization of tool graphs

The system can generate more robust workflows by focusing on structure over surface imitation, as shown by the finding that Evoflux improves weak initial graphs through execution-guided repair, whereas SFT and SFT+DPO often preserve shallow tool-call templates rather than reliable workflow structure.

5. Bold header: Adaptive search intensity based on progress

The system dynamically adjusts its search strategy using an adaptive intensity controller where the growth level determines the search intensity to manage exploration versus exploitation, ensuring that Low growth increases exploratory mutation and may trigger meta guidance.

Small language models (1.5B-4B parameters) offer low-cost, low-latency tool use but often produce plausible plans that fail during execution due to unavailable tools, missing arguments, or broken dependencies. We introduce EvoFlux, an inference-time search method that repairs tool plans through an evolutionary loop over structured graphs of tool calls and dependencies. EvoFlux executes candidate plans against live tool servers and uses execution errors to guide targeted edits, including replacing tools, correcting parameters, inserting missing steps, and removing redundant ones. On held-out MCP-Bench tasks spanning 28 live servers and 250 tools, EvoFlux improves execution success from roughly 3% to 17-24% across four small planners. SFT and SFT+DPO trained on the same 177 demonstrations and 624 preference pairs match or underperform the zero-shot baseline, while ReAct reaches higher peaks on stronger planners at 4-20x greater planner-token cost per solved task. Execution-guided inference-time repair reliably improves small tool-using agents when supervision is limited.