Elias Lumer, Sahil Sen, Vamse Kumar Subbiah, Kevin Paul · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3839851
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
When an agent recurses over a long-context task, should the recursive unit be a bare model call or a full harness with filesystem tools, code execution, and planning? We study this question on long-context aggregation, tasks where a single instance may contain thousands of key--value pairs each requiring individual model reasoning. Coding agents reduce per-entry reasoning to regex heuristics. Recursive language models (RLMs) invoke LLMs recursively but lack tool access. We introduce the Recursive Agent Harness (RAH), in which a parent agent generates and runs an executable script that programmatically spawns full subagent harnesses in parallel for fine-grained workloads and uses structured function calls for subtasks of 1-5 entries. We call this harness recursion and study it against model recursion, recursion over bare model calls as in RLMs, on long-context aggregation. With the backbone held fixed at GPT-5 to match the published Codex and RLM baselines, RAH improves the Codex coding-agent baseline from 71.75% to 81.36% on Oolong-Synthetic (199 samples, 13 context-length buckets up to 4M tokens), a gain attributable to the harness rather than the model. With a stronger backbone, Claude Sonnet 4.5, the same design reaches 89.77%.
arXiv (Cornell University) 2026 · 0 citations
Last synced
This work has 12 recorded citations, but citing papers have not been linked locally yet.
No comments yet — start the discussion below.