Science MapRRLMPrototype
Field scan · 19 Sept 2026 · 150 papers read

Recursion is now being trained. Corpus-level aggregation is not, and that empty row is the opening.

Of the 150 papers read, 16 target global or aggregative queries and not one trains a model for them: eight are benchmarks, eight are prompted pipelines. RL-trained recursion (RAO, context folding, TimeRLM) and a formal corpus-RLM (C-RLM) have all appeared within the last year, so the window for claiming the trained, cost-bounded version is measured in months.

Thesis under test

Recursive Reinforcement Language Models: an RL-trained recursive scaffold that decomposes, retrieves and routes over million-token corpora, matching a read-everything long-context oracle at a small and shrinking fraction of its compute. Core pieces: context-as-environment recursion (RLM), a corpus-privileged adversarial query generator that co-trains a cheap retriever on global/aggregative queries, a learned router over {retriever, symbolic index, oracle}, amortization (index once, answer many), and a taxonomy of mental operations used to diagnose and reward reasoning traces.

The field, decomposed

Area ∝ arXiv volume, last 4 quarters · tiles are the analysed papers, sized and shaded by citations

Tool & Function CallingLong Context, Memory & AmortizationRL, Self-Play & Distillati…Reasoning Traces &…Agentic Search & Dee…Routing, De…Global Co…
Recursive & Programmatic ContextLong Context, Memory & AmortizationGlobal Corpus QA & Structured RAGAgentic Search & Deep ResearchRL, Self-Play & DistillationRouting, Delegation & Compute BudgetsReasoning Traces & Process EvaluationTool & Function Calling
Papers analysed150from 5,161 candidates
On-topic pool3,363used for term trends
From the last 9 months5033% of the analysed set
Report compute cost24%as a result, not a footnote
Evaluate the trace21%not only the final answer

Top bets

Ranked by the opportunity score; full reasoning on the Opportunities tab.

What the literature says

F1

No paper trains for aggregative queries

All 16 papers on global or aggregative QA are either benchmarks (Oolong, CorpusQA, GlobalQA, SummHay, Loong) or untrained pipelines (GraphRAG, S-RAG, LCM, Recursive Agent Harnesses). The SFT, RL, self-play and preference cells of that row are all empty. The best untrained baseline on GlobalQA reaches 6.63 F1.

F2

Others now RL-train recursion, and C-RLM formalises the corpus version

Recursive Agent Optimization (May 2026) RL-trains agents that delegate to recursive copies of themselves. Context folding trains branch-and-fold with FoldGRPO, and TimeRLM's RL post-training cuts turns about 3x. C-RLM proposes offline corpus preparation plus budgeted typed recursive plans with a learned planner, which is close to this agenda, though its abstract shows no experiments. The recursive area's share of AI submissions grew 179% year on year.

F3

Recursion overspends, and the decision to recurse is unowned

An independent RLM reproduction finds that depth 2 hurts accuracy and stretches latency from 3.6 s to 344.5 s. SRLM finds recursion is not RLM's main driver, and a non-recursive admissibility-first build matches RLM on OOLONG-Pairs (0.921 vs 0.906). No analysed paper learns per query whether to recurse, retrieve or read everything.

F4

Cost goes unmeasured where the field grows fastest

Of the analysed routing papers, 79% report compute or cost as a result, against 11% in global corpus QA and 13% in agentic search. Agentic search is the fastest-growing area (+228% share year on year), and its RL agents run trajectories of 100+ turns and 400k tokens with no cost term in the reward.

F5

Self-play needs a verifier, and a corpus can be one

Self-play succeeds where reward is exact: code executors in Absolute Zero, Verus in PSV. Search Self-play validates a question by running RAG over a handful of documents, which cannot verify a count or top-k over a whole corpus. Survive-or-Collapse shows that a ground-truth-aware proposer without a data gate can speed up collapse. No paper grounds a query proposer in programmatic corpus answers.

F6

Reasoning taxonomies exist, but none is used as a reward

Cognitive Foundations labels 192K traces with 28 elements, and the four cognitive behaviours predict RL gains, yet both are used only for diagnosis. Outside the traces area, at most 22% of papers evaluate the trace. FACE-Eval shows that models verbalise least what they learn through tool returns, the very channel through which RLM information arrives. Observable REPL actions are therefore a safer place to reward than chain-of-thought text.

Areas at a glance

Share of AI arXiv submissions matching each area, per 1,000, by quarter.

AreaShare trendYoYCites / moPapers
Recursive & Programmatic Context
179%0.422
Long Context, Memory & Amortization
52%0.416
Global Corpus QA & Structured RAG
68%0.319
Agentic Search & Deep Research
228%0.724
RL, Self-Play & Distillation
90%0.726
Routing, Delegation & Compute Budgets
35%0.614
Reasoning Traces & Process Evaluation
116%0.220
Tool & Function Calling
201%0.39