A New Axis of Scaling

The Kimi k1.5 report reads like a deliberate counter-argument to the entire R1 discourse. Where DeepSeek's paper was terse about methods, Kimi's team published a full training recipe. Where R1 emphasized a novel algorithm (GRPO), Kimi argued the algorithm barely matters — what matters is context length. The paper's thesis: scaling reinforcement learning is a new axis of progress, distinct from pretraining scale, and its key dimension is how long the model is allowed to think.

RL as a new scaling axis

The framing is deliberate and ambitious. Pretraining scales with data, and high-quality data is running out. RL, the authors argue, lets models generate their own training signal by exploring against rewards — a source of improvement that doesn't depend on scraping more of the internet. The goal of k1.5 was to test whether that axis could carry a model to frontier performance on its own.

The recipe has four ingredients, and the paper is explicit that none of them is exotic:

• Long context scaling. RL training context was pushed to 128k tokens, with performance improving steadily as context grew. The engineering trick is partial rollouts: reuse large chunks of previous trajectories instead of regenerating from scratch, which attacks the dominant cost of long-context RL — generation. The paper's framing is that context length is a scaling dimension on par with parameters and data: lengthen the context, get measurably better reasoning, with no new algorithm required. If that curve holds, it reorders intuitions about where capability comes from.

The multimodal results deserve emphasis too: 74.9 on MathVista came from a model jointly trained on text and vision, reasoning across both. Most reasoning-RL work is text-only; k1.5 suggests the RL-discovered reasoning strategies transfer across modalities — which matters if agents, not chatbots, are the endgame.

• Improved policy optimization. A variant of online mirror descent, plus a sampling strategy, a length penalty, and careful data curation. Solid, unglamorous optimization work.
• A deliberately simple framework. No Monte Carlo tree search. No value functions. No process reward models. The paper treats this minimalism as a result: strong performance without the machinery everyone assumed was necessary.
• Multimodality. Joint training on text and vision, with reasoning that spans both.

The pipeline runs pretraining → vanilla SFT → long-CoT SFT → RL, and the RL prompt set is curated with unusual care: diverse coverage, balanced difficulty, and — critically — accurate evaluability. Prompts prone to reward hacking (multiple-choice, true/false, proofs with guessable answers) are filtered out, including a clever test where a model tries to guess answers without reasoning; if it succeeds within 8 attempts, the prompt is discarded as too hackable.

The results that forced attention

Long-CoT k1.5 hit 77.5 on AIME, 96.2 on MATH 500, 94th percentile on Codeforces, and 74.9 on MathVista — matching OpenAI's o1, the closed model that had defined the reasoning frontier. But the more surprising result was the long2short transfer: techniques for distilling long-CoT capabilities into short-CoT models (length penalties plus model merging) produced a short-CoT model scoring 60.8 on AIME and 94.6 on MATH 500 — outperforming GPT-4o and Claude 3.5 Sonnet by up to 550% on these reasoning benchmarks.

That second result is arguably the more important one. It says the reasoning ability RL discovers isn't inherently tied to rambling: the capability can be separated from the verbosity. The short-CoT numbers deserve a second look in this light — 60.8 on AIME from a model that doesn't ramble means strategy and throat-clearing are separable properties. If that separation is robust, the "overthinking" tax on reasoning models is a temporary engineering problem — distill the strategy, discard the rambling — not a permanent cost of intelligence. The paper's open question, reducing overthinking without hurting exploration, is really asking whether that separation can be made clean.

What the simplicity claim conceals

"Simple" here means conceptually simple, not easy. 128k-context RL with partial rollouts is a serious infrastructure achievement, and the prompt-curation pipeline — difficulty estimation via the model's own pass rates, hackability filtering — is doing quiet heavy lifting. A lab without Kimi's infrastructure couldn't reproduce this by reading the algorithm section alone.

The paper is also candid about what's missing. Its conclusion names two open problems: improving the efficiency of long-context RL training, and — notably — "improving credit assignment and reducing overthinking without hurting the model's exploration abilities."

The prompt-curation pipeline is doing more theoretical work than it appears. By filtering for accurate evaluability and balanced difficulty, the team effectively constructs the conditions under which outcome-based RL is well-posed — verifiable rewards, informative gradients, no hacking shortcuts. That's a design pattern worth generalizing: as RL moves to new domains, curating the prompt set may matter as much as the algorithm, because it defines the task distribution the policy actually learns from. That last phrase is doing a lot of work: the field still doesn't know how to make models think less without making them explore less, and the tension between efficiency and exploration may be the central technical problem of reasoning-RL.

Kimi's bet: that the next scaling law isn't about parameters or data — it's about how long a model can sustain coherent thought under RL, and that length is a trainable property.

Where it sits in the arc

Kimi k1.5 is the independent confirmation that made R1 a paradigm rather than a fluke. Two labs, different algorithms (mirror descent vs. GRPO), same conclusion: large-scale RL with verifiable rewards produces reasoning, and you don't need the complex scaffolding. Together with DAPO's open recipe, it completed the 2025 trifecta that turned reasoning-RL from a DeepSeek curiosity into the field's default approach — and its long2short results point at where the paradigm goes next: reasoning that's both powerful and cheap.

Paper: Kimi k1.5: Scaling Reinforcement Learning with LLMs — Kimi Team, 2025.