Alex J. Champandard argues Recursive Language Models lack novelty, debating Grad over whether the paradigm differs fundamentally from Claude

Original post

The authors claim that none of the specific ideas of RLM are novel nor unique to it. So you can't isolate that single concept of "programmatic access to context" alone, it must be assessed more broadly along with the other ideas of RLM — including the persistent REPL and the specifics of the mechanisms used for recursion (both of which also used first elsewhere).

From that perspective, and along those axes, Claude is clearly a different solution unless you weaken the definition of RLM to the point it's meaningless...

Grad@Grad62304977

Tbf a significant part of RLM is programmatic tool calling which Anthropic had blogs on months after the RLM initial blog with pretty big gains (I think they use it for certain evals too) Also things like tool search which many adopted and made blogs on for their importance, is a part of the RLM idea (iirc these were after the initial RLM blog in oct 2025) Also most of the cursor dynamic context discovery blog done around 2 months after the RLM blog are ideas RLM would have

https://claude.com/blog/improved-web-search-with-dynamic-filtering

https://www.anthropic.com/engineering/advanced-tool-use

https://cursor.com/blog/dynamic-context-discovery

1:09 AM · May 30, 2026 · 51 Views

VIEWS64LIKES1REPLIES1

Grad@Grad62304977

I don’t think this makes sense RLM as an idea from my perspective is mainly the combination of subagents, PTC (programmatic tool calling), and programmatic context discovery

If Claude starts saying how PTC is really good for something, it’s not invalid to say it’s more RLM like as it’s something an RLM could do and a key part of its performance

Alex J. Champandard 🌻@alexjc

From that perspective, and along those axes, Claude is clearly a different solution unless you weaken the definition of RLM to the point it's meaningless...

31d6410

Grad@Grad62304977

@alexjc @willdepue @georgejrjrjr No but calling tools like subagents, and reading large contexts through the REPL is a core idea here not an interpretation

Alex J. Champandard 🌻@alexjc

It's certainly moving in interesting directions that are adjacent to RLM.

But the PTC part is your interpretation, the original definitions are focused on *REPL* (persistent language environment) -- which was done before RLM was a twinkle in the authors eyes.

PTC is imho very different than the REPL; having a good API for headless agents is also a new axis vs. tool calls; writing upfront orchestration scripts and not having LLM go step by step also differs.

Lots of subtleties getting lost because of poor definitions. It's all valuable work...

31d3910

Grad@Grad62304977

@alexjc @willdepue @georgejrjrjr They cited them on the idea of programmatic context discovery (actually the part everyone is saying is the most obv) But if u see the blogs I linked, other aspects from RLM can provide big gains and things that Anthropic and cursor made blogs and introduced months after RLMs

30d91

Grad@Grad62304977

- No I meant tools like subagents, can see the blogs I linked too for examples but these are ideas in RLMs

- Well it’s not only the codebase but sure

Mainly getting at ppl not holding cursor and Anthropic to the same standard as they release blogposts on this stuff without citing previous works (not even sure which ones there are other than RLMs) Here referring to all the blogs I linked

30d71

Alex J. Champandard 🌱@alexjc

@Grad62304977 @willdepue @georgejrjrjr - Hmm, there's a section in the paper about the semantics of the subagent calls, you mean calling "subagents like tools" rather? - Codebase (as large context) discovery via tool calls go back before RLM too; it's a core RLM-like concept invented before RLMs.

30d12

Alex J. Champandard 🌱@alexjc

Well, I mean Anthropic did previously cite RLM that's what caused all the credit-assignment arguments to snowball. The current workflows feature imho go a step *against* RLM (as precisely defined): less REPL more upfront-orchestration -- which is interesting concept to discuss if it wasn't for the naming disagreements!

As for the non-technical stuff, the discussion of research as discovery vs. invention is not one that's suitable for this format though ;-)

30d10

Alex J. Champandard 🌱@alexjc

It's certainly moving in interesting directions that are adjacent to RLM.

But the PTC part is your interpretation, the original definitions are focused on *REPL* (persistent language environment) -- which was done before RLM was a twinkle in the authors eyes.

Lots of subtleties getting lost because of poor definitions. It's all valuable work...

31d2

Alex J. Champandard 🌱@alexjc

That's the risk of publishing a paper that collects a variety of existing ideas together (often, no citation should be expected). Different decisions made along the same axes as RLM does not make the solution an RLM, does it?

I think this seems to be the biggest remaining point of disagreement: > The current dynamic workflows of Claude are not a REPL, thus a different approach than recommended by RLM.

Would you agree?

FWIW, I don't think either is better. The upfront workflow scripts more suited to production / heavy processing, and the REPL more suited to exploration / research, but they are conceptually different.

30d1