RLM
A language model training paradigm based on reinforcement learning that uses another model as the source of feedback signals in place of human annotations, aimed at improving training scalability. It belongs to the same language model alignment domain as RLHF.
Timeline (last 90 days)
The trade-off of the RLM approach is that the latency and token overhead of writing and executing code each turn are far greater than direct inference
The RLM Python subprocess is an ordinary subprocess and does not go through an OS-level sandbox (Seatbelt/bwrap); that mechanism only covers shell commands
RLM's constraint mechanisms include: stdout exceeding 1000 characters is converted to a handle, sub-calls are locked to the Flash model, each sub-call has a 120-second timeout, the default max_tokens is 4096, a single RPC batch contains at most 16 prompts, and after recursion depth is exhausted the system falls back to ordinary LLM completion
CodeSmith is an open-source project whose RLM (Recursive Language Model) core design philosophy is to have the large model write code and hand it to an interpreter for execution, rather than having the model perform calculations itself
The RLM system prompt specifies that the model is the root node of a recursive language model, that input is loaded into a long-running Python REPL, and that the model only receives a real-time context handle rather than the raw content
In RLM, input P exists solely as a REPL variable (context/ctx) and never appears in the root LLM's window; the root LLM only receives small-volume metadata messages
The RLM REPL provides helper functions: context_meta() returns restricted metadata, peek(start, end) retrieves a bounded slice by offset, search(pattern, max_hits=100) performs regex search with a hit cap, and sub_query(prompt, slice) delegates semantic judgment to a sub-LLM
The RLM contract specifies that each turn outputs only one repl code block and prose-only turns are not permitted
Alex Zhang is the first author of the RLM paper and a PhD student at MIT
Prime Agent采用递归语言模型（RLM），不提供固定工具集，而是提供持久化的iPython内核，让模型通过编写Python代码创造工具
All Facts (10)
The trade-off of the RLM approach is that the latency and token overhead of writing and executing code each turn are far greater than direct inference
50%UnverifiedThe RLM Python subprocess is an ordinary subprocess and does not go through an OS-level sandbox (Seatbelt/bwrap); that mechanism only covers shell commands
50%UnverifiedRLM's constraint mechanisms include: stdout exceeding 1000 characters is converted to a handle, sub-calls are locked to the Flash model, each sub-call has a 120-second timeout, the default max_tokens is 4096, a single RPC batch contains at most 16 prompts, and after recursion depth is exhausted the system falls back to ordinary LLM completion
50%UnverifiedCodeSmith is an open-source project whose RLM (Recursive Language Model) core design philosophy is to have the large model write code and hand it to an interpreter for execution, rather than having the model perform calculations itself
50%UnverifiedThe RLM system prompt specifies that the model is the root node of a recursive language model, that input is loaded into a long-running Python REPL, and that the model only receives a real-time context handle rather than the raw content
50%UnverifiedIn RLM, input P exists solely as a REPL variable (context/ctx) and never appears in the root LLM's window; the root LLM only receives small-volume metadata messages
50%UnverifiedThe RLM REPL provides helper functions: context_meta() returns restricted metadata, peek(start, end) retrieves a bounded slice by offset, search(pattern, max_hits=100) performs regex search with a hit cap, and sub_query(prompt, slice) delegates semantic judgment to a sub-LLM
50%UnverifiedThe RLM contract specifies that each turn outputs only one repl code block and prose-only turns are not permitted
50%UnverifiedAlex Zhang is the first author of the RLM paper and a PhD student at MIT
50%UnverifiedPrime Agent采用递归语言模型（RLM），不提供固定工具集，而是提供持久化的iPython内核，让模型通过编写Python代码创造工具
50%