Working Notes on Agent Systems/Brad Zhang

@teach_fireworks / X longform

RLMs are the new inference models. Inference models are the first clear demonstration t...

RLMs are the new inference models. Inference models are the first clear demonstration that language model capabilities can be extended with test-time computation. Recursive lang...

April 21, 2026 · 2 min read

RLMs are the new inference models.

Inference models are the first clear demonstration that language model capabilities can be extended with test-time computation.

Recursive language model (RLM) is a new inference model and trend.

This blogger introduced in detail that RLMs inference model can achieve capacity expansion through test-time calculation.

Recursive language models (RLMs) integrate reasoning capabilities and tool invocation into a unified reasoning abstraction paradigm.

The recursive language model treats prompt words as parsable runtime environments, slices them, and initiates recursive sub-calls to break the context length limit.

This model outperforms the current top large language models on long context tasks, long text memory benchmarks such as LongMemEval, and long-range reasoning tasks such as LongCoT.

The latest research results show that recursive language models can convert long inputs into structured calculations, and small models can achieve significant performance improvements with the help of recursive mechanisms.

At present, this technology still has challenges in terms of cost, uncontrollable time-consuming inference, and how to make the model efficient and recursive.

However, the recursive language model points out the development direction of the next generation inference paradigm.

Brief description of terms • Reasoning models: reasoning models • test-time compute: test-time calculation (dynamic expansion of computing power in the reasoning phase) • RLMs: Recursive Language Models, recursive language models • frontier LLMs: frontier large language models • long-horizon reasoning: long-range reasoning • recursion: recursion

Visual summary

Article argument map

Generated from the post's content graph

FORMATTOPICCAPABILITYMARKETcoverscoverscoverscoverscoverssignalsFORMATlongform noteTOPICtechnical distributionTOPICmemoryTOPICinferenceTOPICretrievalTOPICevaluationCAPABILITYevaluation
Mermaid outline
flowchart LR
  format-long_post["longform note"]
  topic-technical-distribution["technical distribution"]
  topic-memory["memory"]
  topic-inference["inference"]
  topic-retrieval["retrieval"]
  topic-evaluation["evaluation"]
  capability-evaluation["evaluation"]
  format-long_post -->|covers| topic-technical-distribution
  format-long_post -->|covers| topic-memory
  format-long_post -->|covers| topic-inference
  format-long_post -->|covers| topic-retrieval
  format-long_post -->|covers| topic-evaluation
  format-long_post -->|signals| capability-evaluation

Visual structure

Essay structure map

Built from summary and key paragraph positions

RLMs are the new inference models. Inference models are the first clear demonstration...THESISRLMs are the newinference models.Inference models arethe first clearSIGNALRLMs are the newinference models.OPERATORThe recursive languagemodel treats promptwords as parsableruntime environments,IMPLICATIONBrief description ofterms • Reasoningmodels: reasoningmodels • test-time
Mermaid outline
flowchart LR
  thesis["RLMs are the new inference models. Inference models are the first clear demonstration that language model c..."]
  signal["RLMs are the new inference models."]
  operator["The recursive language model treats prompt words as parsable runtime environments, slices them, and initiat..."]
  implication["Brief description of terms • Reasoning models: reasoning models • test-time compute: test-time calculation..."]
  thesis -->|frames| signal
  signal -->|develops| operator
  operator -->|lands in| implication

Source: View the original post