Editorial illustration for New self-summarization RL technique trims action history to 1,000 tokens
RL Models Trim Memory with Smart Token Reduction
New self-summarization RL technique trims action history to 1,000 tokens
MiniMax just dropped a training trick that makes you wonder why nobody thought of it sooner. Their new self-summarization RL technique forces the model to hit pause when its internal token count runs long, compress its own action history down from 5,000-plus tokens to roughly 1,000, and keep earning rewards across the entire trajectory. Cursor, which has been quietly training its Composer 2 on top of MiniMax’s Kimi, reports 50% fewer compaction errors and measurably stronger long-horizon task handling.
The broader story here isn’t just a clever optimization. Chinese labs like MiniMax keep delivering models that can now handle what only Western closed-source players used to own. And for anyone who scoffed at Cursor for building on open foundations, that take is aging fast.
Starting with a strong open model and training it further isn’t a shortcut; it’s the obvious next move.
A key technical novelty is "self-summarization," a compaction-in-the-loop RL method that trains the model to pause on token-length triggers and compress its own action history to ~1,000 tokens from 5,000+, with rewards spanning the entire trajectory; Cursor reports 50% fewer compaction errors and stronger long-horizon task handling. Editor's Take: MiniMax and Chinese labs in general continue to impress with their even improving models, which at this point are more than capable of taking care of a lot of stuff that only Western closed source models used to be capable of. Cursor got some flak for training Composer 2 on top of Moonshot AI's Kimi, which is quite silly - starting with already strong open source models and training them further should by now be the no-brainer move for any AI company whose primary business isn't already frontier model development.
This is not a footnote in the AI arms race; it’s a quiet but decisive shift in the calculus of practical deployment. Self-summarization isn’t a flashy benchmark killer, it’s a workhorse that lets agents hold a coherent conversation across a thousand steps without drowning in their own exhaust. MiniMax and peers like them are proving that the real frontier isn’t just bigger models; it’s smarter loops.
Cursor’s 50% crash in compaction errors is the sort of metric that signals a threshold: when open-source foundations can be fine-tuned into tools that just *work* on long-horizon tasks, the moat around proprietary systems starts to look like a puddle. The snickering over training on Kimi? Misplaced.
That’s how winning looks now, take strong, make it stronger, and ship. The action history is trimmed. The conversation should follow.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Human-in-the-Loop: Training Wheel Mode Lets Agents Prove Themselves in Risky Ops
From Archives Filmmaker revisits generative AI white‑paper authors, sees eugenic undertones
Common Questions Answered
How does the new self-summarization technique reduce action history length in reinforcement learning?
The technique trains the model to pause at predefined token-length triggers and compress its own action history from over 5,000 tokens to approximately 1,000 tokens. This approach allows the model to dynamically manage its context window, preventing excessive computational overhead and maintaining long-horizon task performance.
What performance improvements does the self-summarization method offer compared to traditional approaches?
According to the report, the self-summarization method reduces compaction errors by 50% while maintaining strong performance across long-horizon tasks. The technique rewards the entire trajectory, ensuring that the model's learning and performance are not compromised during the token compression process.
Why is managing token length critical in reinforcement learning language models?
In reinforcement learning setups, action logs can quickly expand beyond five thousand tokens, which pushes the system toward context-window limits and necessitates costly recomputation. By implementing a dynamic self-summarization approach, models can efficiently manage their memory and maintain computational efficiency without losing critical contextual information.
Further Reading
- Self-Hinting Language Models Enhance Reinforcement Learning — arXiv
- Self-Hinting Language Models Enhance Reinforcement Learning — Hugging Face Papers
- Reinforcement Learning via Self-Distillation (Jan 2026) — YouTube
- The State Of LLMs 2025: Progress, Problems, and Predictions — Sebastian Raschka's Magazine
