Analysis, commentary, and technical deep dives on Recursive Language Models and the shifting landscape of AI architecture.
How to put an RLM behind an MCP tool or give its REPL MCP tools, and why MCP sampling is the wrong base for sub-calls.
October 2026The reference library caps the root loop at 30 turns and DSPy at 20. Here is what the cap does, what happens when a run hits it, and how to set it.
October 2026The reference library already runs batched sub-calls with asyncio.gather and a semaphore. Here is how it works and when to write your own.
October 2026Accuracy against reasoning length is an inverted U. The evidence, the vendor guidance, and the rules for RLM thinking budgets and recursion depth.
October 2026No official trajectory dataset exists, but the weights, the logger and the paper's filtering recipe are public. What is out there and how to build your own.
September 2026The KV cache formula with worked 128K numbers for Llama 3.1, Mistral NeMo and Phi-3, plus GQA, FP8 and 2-bit savings.
September 2026A source-level comparison of dspy.RLM and the reference rlms library: recursion, budgets, sandboxes and training.
September 2026Why small models loop inside llm_query, what the degeneration research says, and the settings that fix it.
September 2026What the RLM code accepts, what RVLM built, and how to add vision sub-calls yourself.
September 2026What the rlm library, the RLM papers, and the provider docs actually say about splitting temperature between the root loop and its sub-calls.
September 2026Why RLM root models invent execution results, what the official repo changed, and the stop-sequence and tool-call fixes.
September 2026How provider and local prefix caches interact with RLM root turns and sub-calls, and how to structure prompts to save money.
September 2026What the reference RLM library's local REPL blocks, why prompt injection makes isolation mandatory, and a production sandbox policy.
September 2026Mechanics of multi-hop QA with an RLM: corpus as a REPL variable, regex filters, one sub-call per candidate, and the BrowseComp-Plus results at 1,000 documents.
September 2026Same tool loop, different home for the context. Where an RLM and a shell-equipped coding agent overlap, where they split, and the Table 1 numbers.
September 2026What compaction keeps and loses, what an RLM keeps that a summary cannot, and the hybrid that uses both.
September 2026Yes -- 3.6 seconds became 89.3 at depth 1 in one reproduction. But the slowdown lives in sequential, blocking sub-calls, not the paradigm, and the median RLM run is actually cheaper. Where the time goes and when to accept it.
August 2026Not a huge context window. The root model needs reliable code emission in a REPL, faithful instruction following, and output-token headroom after thinking. Which models clear the bar, and how 1,000 trajectories got an 8B model there.
August 2026Depth 1 captured most of the gain at flat or lower cost. Depth 2 and 3 paid off only on quadratic aggregation with a frontier coder, and hurt weak coders and simple retrieval. The evidence at each depth and a decision rule.
August 2026Accuracy alone tells you almost nothing. The complexity class of the benchmark, the depth-0 ablation, the latency tail, and the answer-extraction failure rate are what decide the number.
August 2026The documented failure modes: base models that cannot write correct code, recursion depth that only pays off sometimes, brittle answer detection, and a long cost tail.
February 2026Everyone is writing about self-improvement loops. The real unlock is when one model teaches another -- not its outputs, but its learning process.
February 2026Voltropy's new paper doesn't compete with RLM. It builds on it, proves it works, and takes it further than we did.
February 2026The industry spent five years making context windows bigger. RLMs suggest the entire framing was wrong from the start.
February 2026A practitioner's guide to the core algorithmic pattern behind Recursive Language Models -- and why it mirrors how expert humans solve complex problems.
February 2026Retrieval-Augmented Generation dominated the long-context conversation for three years. RLMs take a fundamentally different path. Here is where each one wins.
February 2026Training-time scaling dominated the last decade of AI progress. The next decade belongs to inference-time scaling -- and RLMs are a leading example of why.
February 2026From legal discovery to genomics, the tasks where RLMs create the largest performance gap over standard LLMs.
February 2026How an 8B parameter model post-trained on 1,000 samples can rival GPT-5 on long-context tasks -- and what that means for cost and deployment.