Recursive Language Models

2025/12/31 by Alex L. Zhang, Tim Kraska, Omar Khattab · 40 voices · 4 citations
#cs.AI #cs.CL

paper · pdf

Abstract

We study allowing large language models (LLMs) to process arbitrarily long prompts through the lens of inference-time scaling. We propose Recursive Language Models (RLMs), a general inference paradigm that treats long prompts as part of an external environment and allows the LLM to programmatically examine, decompose, and recursively call itself over snippets of the prompt. We find that RLMs can successfully process inputs up to two orders of magnitude beyond model context windows and, even for shorter prompts, dramatically outperform the quality of vanilla frontier LLMs and common long-context and coding scaffolds (e.g., on GPT-5 by a median across the evaluated benchmarks of 26% against compaction, 130% against CodeAct with sub-calls, and 13% against Claude Code) across four diverse long-context tasks while having comparable cost. At a small scale, we post-train the first model around the RLM. Our model, RLM-Qwen3-8B, outperforms the underlying Qwen3-8B model by 28.3% on average and even approaches the quality of vanilla GPT-5 on three long-context tasks. Code is available at https://github.com/alexzhang13/rlm.

Citations

• Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention
• Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking
• Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation
• YOLOv3: An Incremental Improvement

Cited by

• Systematic Literature Reviews With Two Multi-Agentic Systems And Human-In-The-Loop
• Continual Harness: Online Adaptation for Self-Improving Foundation Agents
• Eta Given Delta: Defining LLM Tool Efficiency With Marginal Tool Utility
• LCM: Lossless Context Management

Discussions

• Recursive Language Models [hn, 161 points, 23 comments]
• MIT just published a paper that demos how to fix "context rot" by using using recursion. They chunk up and offload tasks to smaller child LLMs. Apparently response quality for large contexts doesn't t [bsky, 9 points, 1 comments]
• arxiv.org/abs/2512.24601 [bsky, 7 points, 1 comments]
• Recursive Language Models View Article | Join the HN Conversation Summary of HN discussion 🧵👇 [bsky, 3 points, 1 comments]
• Sure. This is mostly true of older LLM architectures, but they're not the only thing out there. I'm seeing plenty of evidence in the field that not only can newer architectures do symbolic reasoning, [bsky, 3 points, 1 comments]
• Larger context windows make this problem *worse* not better. It shows that there is substantial degradation after as little as 16k tokens. Here's MITs data from six months ago arxiv.org/pdf/2512.24601 [bsky, 2 points, 0 comments]
• What's the best way to expand the context window of a language models? With more language models! MIT just demonstrated a technique for chunking a model's context to allow those chunks to be processed [bsky, 2 points, 0 comments]
• Okay, so this makes me want to keep researching AI. It basically introduces symbolic representations and relies on the LLM to construct the program/language. arxiv.org/abs/2512.24601 [bsky, 1 points, 0 comments]
• arxiv.org/abs/2512.24601 I've been implementing something similar to this paper, against 8b models locally and it's been really interesting watching their capability grow from the tiniest little thing [bsky, 1 points, 1 comments]
• I took a shot at implementing the idea from the Recursive Language Model paper [lemmy, 1 points, 0 comments]
• https://bsky.app/profile/buzzing.cc.web.brid.gy/post/3mbkpsyvd7yf2 [bsky, 1 points, 0 comments]
• In der AI-Bubble auf linkedin schreibt ein Experte: Ich habe den Super-Prompt entwickelt, aufgrund eines Scientific Paper des MIT. Schreibe „XYZ“ in die Kommentare und du erhältst den link. Mehr als 3 [bsky, 1 points, 0 comments]
• この論文めっちゃ面白い😍 長い文章をLLMで正確に扱うには？→REPLに変数として保持し、コードを通じてLLMがそれを操作する。LLMは長文を直接参照せず、どこを見るべきかの仮説を立ててコードを生成・実行することで、必要な情報を取得し回答を生成する arxiv.org/abs/2512.24601 [bsky, 1 points, 0 comments]
• [arXiv] Recursive Language Models Summary Recursive Language Models (RLMs) propose an inference-time strategy to overcome LLM context limits by treating long prompts as external data. arxiv.org/pdf/25 [bsky, 1 points, 0 comments]
• Recursive Language Models (RLM): treat memory as an external environment + recursive calls. Interesting direction for long-context agents. Paper: https://arxiv.org/abs/2512.24601 Code: https://github. [bsky, 0 points, 0 comments]
• RLMs treat the prompt as data inside a REPL and let the LM recurse on snippets—handling 10M+ tokens and beating long-context baselines on tough tasks. Simple idea, big wins. Paper: arxiv.org/abs/2512. [bsky, 0 points, 0 comments]
• ⚡ Hackernews Top story: Recursive Language Models [bsky, 0 points, 0 comments]
• [2512.24601] Recursive Language Models [bsky, 0 points, 0 comments]
• With the launch of dspy's RLM() module I took some time to look at the RLM paper, and it's a banger arxiv.org/abs/2512.24601 [bsky, 0 points, 1 comments]
• En train de faire une note de lecture sur les RLM (R pour Recursive). Je ne sais pas encore si c'est du bidouillage où s'il y a quelque chose de plus profond. Papier sur Arxiv (du MIT ça semble crédib [bsky, 0 points, 0 comments]
• It’s all about how you ask the question - promising developments in having LLMs act like real experts. Break problems down and challenge answers arxiv.org/pdf/2512.24601 [bsky, 0 points, 0 comments]
• Successfully plugged #cerebras #gptoss into the framework from the Recursive Language Models paper @ https://arxiv.org/abs/2512.24601 Getting ready to test it out on some long context problems. I've [bsky, 0 points, 1 comments]
• arxiv.org/abs/2512.24601 [bsky, 0 points, 0 comments]
• New version of paper: arxiv.org/pdf/2512.24601 [bsky, 0 points, 0 comments]
• #AI #LongContext #MITCSAIL #mlsky MIT CSAIL proposes Recursive Language Models that treat long prompts as an external environment. The model writes Python3 to inspect, decompose, and recursively query [bsky, 0 points, 0 comments]
• Recursive Language Models https:// arxiv.org/abs/2512.24601 # arxiv [mastodon, 0 points, 0 comments]
• 아니 뭐 이거 괜찮아 보인다고 그래서 봤더니, 그냥 모든 프롬프트를 LLM에 넣는게 아니고, LLM이 prompt를 코드를 통해 읽어서, 작업 분배하고, 재귀적으로 호출하게끔 함. 딱히 LLM이 특화된 것도 아니고, 프롬프팅 기법이랑 뭐가 다른지 모르겠음… 그렇다고 sub-prompt들이 서로 연관되고, 이를 선택하는 똑똑한 알고리즘이 있는 것도 아님. � [bsky, 0 points, 0 comments]
• Recursive Language Models https://arxiv.org/abs/2512.24601 (https://news.ycombinator.com/item?id=46475395) [bsky, 0 points, 0 comments]
• Recursive Language Models https://arxiv.org/abs/2512.24601 (https://news.ycombinator.com/item?id=46475395) [bsky, 0 points, 0 comments]
• Recursive Language Models arxiv.org/abs/2512.24601 [bsky, 0 points, 0 comments]
• Source: arxiv.org/abs/2512.24601 [bsky, 0 points, 0 comments]
• The paper is by Alex Zhang, Tim Kraska, and Omar Khattab at MIT CSAIL. Zhang first introduced RLMs in a blog post back in October 2025. The full paper just dropped with expanded experiments. Paper: ar [bsky, 0 points, 0 comments]
• arxiv.org/pdf/2512.24601 [bsky, 0 points, 0 comments]
• #NoNewWeights Paper: https://arxiv.org/abs/2512.24601 [bsky, 0 points, 0 comments]
• Recursive Language Models [bsky, 0 points, 0 comments]
• Recursive Language Models #HackerNews https://arxiv.org/abs/2512.24601 [bsky, 0 points, 0 comments]
• arxiv.org/abs/2512.246... [bsky, 0 points, 1 comments]
• Recursive Language Models https://arxiv.org/abs/2512.24601 https://news.ycombinator.com/item?id=46475395 [bsky, 0 points, 0 comments]
• Recursive Language Models https://arxiv.org/abs/2512.24601 [bsky, 0 points, 0 comments]
• Recursive Language Models https://arxiv.org/abs/2512.24601 [comments] [44 points] [bsky, 0 points, 0 comments]

Related

• PromptIntern: Saving Inference Costs by Internalizing Recurrent Prompt during Large Language Model Fine-tuning
• Recursive Inference Scaling: A Winning Path to Scalable Inference in Language and Multimodal Systems
• MatryoshkaThinking: Recursive Test-Time Scaling Enables Efficient Reasoning
• Large Language Models Can Self-Improve in Long-context Reasoning
• Brevity is the soul of sustainability: Characterizing LLM response lengths