FlowEvo, a framework accepted at COLM 2026, compiles an agent's successful multi-step workflows into callable executable functions and stores them in a bank that grows during use, with no weight updates at all. On ALFWorld, a benchmark of household tasks, it reached 85.6 percent -- 26.4 points above the strongest of eight baselines, using roughly a third as many tokens. The interesting part is the deletion: it tracks each stored skill's downstream usefulness and suppresses the ones that make later tasks worse.
Key facts
- 85.6 percent on ALFWorld, reported as 26.4 points above the strongest baseline, at roughly one third the tokens.
- Beat the ExpeL baseline in 49 of 50 model-dataset comparisons, across 10 base models from 7B to 671B parameters.
- Training-free: workflows and skills co-evolve at inference time on a shared GPT-4o-mini backbone.
- arXiv 2607.21596, revised August 20, 2026; code at github.com/DEFENSE-SEU/FlowEvo.
There is a specific waste in how agents work today, and once you see it you cannot unsee it. An agent given a complicated task will figure out a procedure -- check this, then that, handle this edge case, verify, report. It executes. It succeeds. And then the procedure evaporates, because it lived in a context window that gets cleared. Tomorrow the same agent solves the same problem from scratch, paying full price again.
The obvious fix is a skill library, and people have built those. The problem, as the authors put it, is that existing libraries "are typically assembled offline and do not grow from the agent's own workflows." Someone writes the reusable routines by hand in advance. That works until the tasks drift away from what the author imagined.
FlowEvo closes the loop. When a workflow succeeds, it gets compiled into a callable skill and written to a persistent bank. Later tasks retrieve it either by executing it directly or by using it as structured context while constructing a new workflow. The bank grows out of the agent's own history rather than a designer's forecast. Ground Truth's explainers on agent memory and in-context learning cover the two mechanisms this sits between.
The analogy is a working cook's recipe box, and the useful half of the analogy is the part people skip. Anyone can add cards. What makes a recipe box good over ten years is throwing cards out -- the technique that worked once under conditions that never recurred, the shortcut that quietly ruins the dish when the ingredients change. FlowEvo does this explicitly: it "tracks each skill's downstream utility and suppresses skills that cause negative transfer." Negative transfer is the failure mode where reusing something learned earlier actively hurts on a new task, and it is the reason naive memory systems get worse as they get bigger.
The results span more than one benchmark, which matters for a claim like this. Using a shared GPT-4o-mini backbone, FlowEvo took the highest accuracy among eight baselines on the full standard splits of ALFWorld, HumanEval, MBPP, GSM8K, and MATH-500 -- household tasks, two code benchmarks, grade-school math, and competition math. That range is the point: a skill bank that only helps on one task family is a specialized cache, not a general mechanism. Across 10 base models from 7 billion to 671 billion parameters, it outperformed the ExpeL baseline in 49 of 50 model-dataset comparisons, which is the kind of near-sweep that usually indicates the mechanism is doing something structural rather than fitting one setup.
The token reduction deserves its own attention. Roughly a third of the tokens on ALFWorld is not a rounding improvement, and it comes from the same source as the accuracy gain -- calling a compiled function rather than re-deriving a procedure in natural language every time. Given that output tokens are the expensive half of every model bill, a method that raises accuracy while cutting generation is unusual enough to be worth checking carefully.
Why it matters: this belongs to a visible shift in where researchers put the learning. Ground Truth has covered several versions of it -- three papers arguing for building the training world rather than the model, and a frozen model that can look like it taught itself. FlowEvo is the most operationally concrete of the family. The memory is executable code in a bank, not prose in a prompt, so you can read it, test it, and delete it. That auditability is the difference between a research demo and something an operations team would agree to run. It also lands the same week that OpenAI open-sourced the harness layer, which is the natural home for exactly this kind of persistent skill bank -- see agent harnesses and scaffolding.
The honest caveats. These are the authors' own reported results, not independent reproductions, and the benchmarks -- ALFWorld in particular -- are well-worn enough that strong numbers on them do not automatically transfer to messy production work. The public repository is real but small, with under two dozen stars at the time of writing, so the community has not yet stress-tested it. And the harder question the paper does not resolve is what happens over a much longer horizon: a skill bank that keeps growing eventually faces a retrieval problem of its own, and suppression by downstream utility is a heuristic, not a guarantee.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)