SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2604.14877 · 2026

Does RL Expand the Capability Boundary of LLM Agents? A PASS@(k, T ) Analysis *

Xin Wang, Wenjing Yan, Zhiyuan Zhai, Xiaodan Shao

arXiv · PDF · Open in the Atlas

Code that ran

2 of 2 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 1 failure of our instrument; 0 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2604.14877.json · badge: /api/ran/2604.14877.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
zhiyuanZhai20/pass-kt-analysis community
Papers with Code archive
2 ran · 0 no recorded run
FunctionStatusWhere it lives
compute_pass_kt_grid Ran
ran; our contract draft was wrong, not the code
zhiyuanZhai20/pass-kt-analysis/eval/compute_pass_kt.py
pointer only (licence: NONE) · get_code("35264e78fb11d3d4")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, zhiyuanZhai20/pass-kt-analysis when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

pass_at_k Ran
ran, honoured the contract we drafted
zhiyuanZhai20/pass-kt-analysis/eval/compute_pass_kt.py
pointer only (licence: NONE) · get_code("340437ebc52f7a3a")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, zhiyuanZhai20/pass-kt-analysis when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

Also present in other papers' code

Of 2 samples lifted from this paper's linked repositories, 2 carry a function identity and 0 do not. Of the 2 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 2 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Does reinforcement learning genuinely expand what LLM agents can do, or does it merely make them more reliable at what they could already do? For static mathematical reasoning, a recent line of work answers the second: base and RL pass@k curves converge at large k, and RL only redistributes probability mass within the base model's existing capability set. We ask whether this pessimistic reading survives the move to agentic tool use, where an agent's T rounds of environment interaction enable compositional strategies that re-sampling cannot recover. We introduce PASS@(k, T ), a two-dimensional evaluation metric that jointly varies a sampling budget k and an interaction-depth budget T , cleanly separating capability expansion from efficiency improvement. Our main finding is that, contrary to the static-reasoning result, tool-use RL genuinely enlarges the agent's capability boundary: as the sampling budget grows, the RL agent's passcurve pulls above the base model's and the gap widens at the right tail rather than converging, the opposite of what the static-reasoning RLVR literature reports. The expansion is specific to tasks that require compositional, sequential information gathering; on simpler categories RL behaves as prior work predicts. Under matched training data, supervised fine-tuning produces the opposite effect: it regresses the capability boundary on the same compositional tasks, which isolates selfdirected exploration, not data exposure, as the causal factor. A mechanism analysis explains how the expansion arises: RL reweights the base model's existing strategy distribution toward the subset whose downstream reasoning more often yields a correct answer, with the improvement concentrated on how the agent integrates retrieved information rather than on what it searches for. These results reconcile the optimistic and pessimistic readings of RL for LLMs: both are correct, on different task types. In agentic settings, when the environment admits compositional solutions the base model cannot stitch together alone, RL meaningfully teaches new capabilities. * Code available at https://github.com/zhiyuanZhai20/pass-kt-analysis.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2604.14877")
get_code_for_paper("2604.14877")
have("2604.14877")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2604.14877.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2604.14877.svg)](https://syntology.ai/paper/2604.14877)

Connect an agent — have() is free.