---
title: 'KAT: Adaptive Reasoning & Agentic Coding'
url: https://www.emergentmind.com/topics/kwaipilot-autothink-kat
type: topic
---

# KAT: Adaptive Reasoning & Agentic Coding

Kwaipilot-AutoThink, usually abbreviated **KAT**, denotes a Kuaishou-developed model family and training paradigm centered on **adaptive reasoning control**: the model is trained to switch between **reasoning mode** and **non-reasoning mode** according to task complexity rather than emitting long chain-of-thought traces indiscriminately [2507.08297]. In its initial public form, KAT is presented as an open-source **40B** dense large language model, **KAT-V1-40B**, aimed at the **overthinking** problem in reasoning-intensive tasks; in later reports, the same KAT line expands into **agentic coding** systems, notably **KAT-Coder**, the open-sourced **32B** checkpoint **KAT-Dev**, and the continued-post-training successor **KAT-Coder-V2** [2507.08297] [2510.18779] [2603.27703].

## 1. Definition, scope, and naming

KAT is defined in the KAT-V1 report as **Kwaipilot-AutoThink**, and in the KAT-Coder report as the **broader model line and training philosophy** under which code-specialized descendants are organized [2507.08297] [2510.18779]. The naming is hierarchical rather than flat: KAT denotes the family-level idea of adaptive or autonomous reasoning, KAT-Coder denotes a code-specialized agentic model, and KAT-Dev denotes the released deployment-oriented checkpoint.

| Name | Meaning | Role |
|---|---|---|
| **KAT** | **Kwaipilot-AutoThink** | Broader model line and training philosophy |
| **KAT-V1-40B** | Dense 40B model | Initial open-source adaptive reasoning model |
| **KAT-Coder** | Code-specialized KAT model | Agentic coding system in the technical report |
| **KAT-Dev** | Open-sourced 32B checkpoint | Deployment-oriented member of the KAT-Coder series |
| **KAT-Coder-V2** | Continued-post-training successor | “Specialize-then-Unify” agentic coding model |

The family is framed not as a generic chat model line but as a response to two related system-level mismatches. In KAT-V1, the mismatch is between reasoning-heavy language models and interactive applications where indiscriminate chain-of-thought causes unnecessary latency and token inflation. In KAT-Coder, the mismatch is between static text-based code-model training and **dynamic real-world agentic execution**, where a model must plan, act, use tools, recover from failures, and operate inside interactive software workflows [2507.08297] [2510.18779].

The reports also place KAT inside a production setting rather than a purely benchmark-driven one. KAT-V1 is described as deployed in **Kwaipilot**, Kuaishou’s internal coding assistant, and KAT-Coder is explicitly designed for **production-grade IDE environments** rather than only offline code completion [2507.08297] [2510.18779].

## 2. AutoThink as adaptive reasoning control

The conceptual core of KAT is the **AutoThink** paradigm: the model should decide whether to enter a deliberative reasoning regime or answer directly, rather than treating explicit reasoning as uniformly beneficial [2507.08297]. The target failure mode is **overthinking**, defined as applying chain-of-thought-style reasoning to tasks that do not require it, thereby increasing latency, inflating token usage, and degrading efficiency in interactive systems.

In the KAT-V1 framing, the two operational modes are **reasoning mode** and **non-reasoning mode**. Reasoning mode corresponds to explicit chain-of-thought generation; non-reasoning mode corresponds to direct answering without multi-step deliberation. The claimed benefits are fourfold: higher or competitive **accuracy** on difficult reasoning tasks, improved **efficiency** through reduced unnecessary reasoning, stronger **controllability**, and substantial **token reduction**, including repeated claims of savings up to roughly **30%** relative to strong reasoning baselines [2507.08297].

The model is designed to support both **autonomous mode switching** and **user-guided mode switching**. Autonomous switching is learned within the autoregressive policy itself: at inference time the model first produces a **`<judge>`** analysis and then emits either **`<think_on>`** or **`<think_off>`**, after which generation proceeds accordingly. User-guided switching remains possible through prompt-level directives such as **“DO NOT THINK”** or **“Think deeper”**, which bias or override the default behavior [2507.08297].

This mode-selection view distinguishes KAT from systems that only improve reasoning quality once reasoning has already been invoked. Related later work on adaptive reasoning is conceptually adjacent but methodologically different. **ThinkPilot** performs **training-free** optimization of `<think> ... </think>` prefixes by behavior-guided evolutionary search, while **Omni-AutoThink** applies a two-stage supervised-plus-RL recipe to multimodal omni models and uses a think/no-think reward asymmetry to induce adaptive reasoning behavior [2510.12063] [2512.03783]. A plausible implication is that KAT belongs to a broader 2025–2026 research shift in which “whether to think” becomes a separate optimization target from “how to think.”

## 3. KAT-V1-40B: model, data, and knowledge transfer

The concrete initial KAT release is **KAT-V1-40B**, a **dense** 40B-parameter model obtained by **upscaling Qwen2.5-32B to 40B** through targeted layer-wise expansion after **layer saturation analysis**. The paper reports **80 layers**, **no tied embedding**, **40 / 8 attention heads (Q / KV)**, and **32K context length**. It explicitly notes that “pre-training” means **continued pre-training based on the upscaled Qwen2.5-32B to a 40B scale**, not training from scratch [2507.08297].

The KAT-V1 training system is a **two-stage framework**. The first stage is described as **pre-training: knowledge enhancement**, where the model learns to follow externally supplied mode cues using explicit trigger tokens **`<think_on>`** and **`<think_off>`**. These tokens are **not involved in loss computation**, but function as mode selectors during training. At this stage the model learns **passive switching** rather than autonomous gating [2507.08297].

A central contribution of this stage is the construction of a **dual-regime dataset**. Queries are first assessed by a **novel data labeling system** using state-of-the-art LLMs to evaluate **difficulty** and **domain characteristics**. **Think-off** queries are those judged not to require multi-step reasoning; for these, **DeepSeek-V3** is used to generate responses, followed by **reject sampling** with multiple state-of-the-art LLMs so that only accurate and relevant completions are retained. **Think-on** queries are processed through a **multi-agent framework** with three roles—**solver**, **thinker**, and **critic**—to generate verified long-CoT traces. The Stage 1 corpus contains **approximately 10 million examples**, with **34.8% Think-on** and **65.2% Think-off** [2507.08297].

Knowledge transfer in Stage 1 uses **MTP-enhanced knowledge distillation**. The teachers are regime-specific: **DeepSeek-V3** for **Think-off** data and **DeepSeek-R1-0528** for **Think-on** data. The student is augmented with **Multi-Token Prediction (MTP)** modules that predict multiple future tokens in parallel, and the report states that combining MTP with a **Universal Logit Distillation Loss** yields **approximately a 5 percent improvement** over standard KD on several benchmarks [2507.08297]. The paper does not provide the exact ULD or MTP objective equations, but it is explicit that token-level logit alignment outperformed the other intermediate alignment strategies examined.

The report also describes a scaling extension beyond the dense 40B checkpoint. An early **KAT-V1-200B** row is reported for a **200B Mixture-of-Experts (MoE)** model with **40B activated parameters**, and the paper uses those early results to argue that the AutoThink paradigm scales beyond the initial dense release [2507.08297].

## 4. Cold start, structured supervision, and Step-SRPO

The second KAT-V1 stage is the **post-training: AutoThink** stage, where externally guided switching is converted into **autonomous mode selection**. This transition begins with a **cold-start initialization** procedure. For each user query, the system uses a **majority voting mechanism across multiple model outputs** to decide the preferred reasoning mode, then samples responses from **DeepSeek-R1** for **Think-On** and **DeepSeek-V3** for **Think-Off**. **DeepSeek-V3** is additionally used to produce a **judgment rationale** explaining why the chosen mode is appropriate; these rationales act as **mode-selection priors** [2507.08297].

An exploration mechanism is built into this cold-start process: **1% of the queries are randomly assigned a mode**, regardless of majority voting. The report states that this is intended to expose the model to diverse reasoning scenarios and reduce overfitting to biased gating patterns [2507.08297].

The Stage 2 supervision format uses explicit structural tokens:

```text
<think-on format>
<judge> {judge_analysis} </judge>
<think_on> <think> {thinking_content} </think>
<answer> {response} </answer>

<think-off format>
<judge> {judge_analysis} </judge>
<think_off> <answer> {response} </answer>
```

This decomposition makes three sub-behaviors explicit: **judgment**, **mode selection**, and **answer generation**. The Stage 2 corpus contains **approximately 3.5 million examples**, with Think-on to Think-off at roughly **2:1**, spanning **code, math, science, general, multi-turn, and tool use** [2507.08297].

The final reinforcement-learning stage is **Step-SRPO**, described as an extension that introduces **intermediate supervision** into a **GRPO/SRPO-style** framework. The paper does not provide the full derivation, but it repeatedly characterizes the method as jointly optimizing **Thinking Control**, **Reasoning RL**, and **General RL** in one training session. The reward is split into **Judge Reward**, based on correctness of reasoning activation, and **Answer Reward**, based on final answer correctness or quality and **modulated by the Judge Reward** [2507.08297].

Reward definitions are given by domain rather than by a single closed-form global equation. In **math**, evaluation uses symbolic equivalence checks; in **code**, predefined unit tests; in **science**, comparison of the first extracted multiple-choice answer letter; in **general**, lightweight heuristics such as keyword overlap and a simple classifier. The Stage 3 RL corpus contains **45K samples** and is biased toward harder examples while retaining a small proportion of easy examples for stabilization [2507.08297].

At inference time, switching remains explicit rather than latent. The trained model first produces a **`<judge>`** analysis, then emits either **`<think_on>`** or **`<think_off>`**. This suggests that KAT operationalizes adaptive reasoning as a structured token-level behavior rather than a separate hidden controller.

## 5. Agentic coding evolution: KAT-Coder, KAT-Dev, and KAT-Coder-V2

The KAT family later broadens from adaptive reasoning in general LLMs to **agentic coding**. The **KAT-Coder Technical Report** describes a large-scale coding model meant to operate in **interactive software development workflows**, and presents a four-stage curriculum: **Mid-Term Training**, **Supervised Fine-Tuning (SFT)**, **Reinforcement Fine-Tuning (RFT)**, and **Reinforcement-to-Deployment Adaptation** [2510.18779].

Mid-Term Training uses a heterogeneous corpus that includes a **real-world software engineering corpus of about 20B tokens** mined from GitHub, reasoning and reflection data, synthetic **Plan–Action–Observation** trajectories, and instruction-following data with logical and structural constraints. The SFT stage constructs **over one million samples**, deliberately balanced across **over twenty mainstream programming languages**, **ten representative contexts**, and **ten task archetypes**. The RFT stage introduces **multiple trajectory ground truths** and a **relative evaluation framework** under **GRPO**, while the deployment-adaptation stage introduces **Error-Masked SFT (EM-SFT)** and **Tree-Structured Trajectory Training (TST)** for production IDE traces [2510.18779].

The coding branch has its own internal naming hierarchy. **KAT-Coder** is the code-specialized technical system; **KAT-Dev** is the open-sourced **32B** deployment-oriented checkpoint hosted at **Kwaipilot/KAT-Dev**; and the family is described as a deployable foundation for real-world intelligent coding agents [2510.18779]. A key systems contribution in this report is **Trie Packed Training**, which treats agent trajectories with shared prefixes as tries and computes those shared prefixes only once, using custom kernels for shared-prefix mask attention and modified positional embeddings.

The subsequent **KAT-Coder-V2 Technical Report** recasts agentic coding under a **“Specialize-then-Unify”** paradigm. Agentic coding is decomposed into five expert domains—**SWE**, **WebCoding**, **Terminal**, **WebSearch**, and **General**—each trained independently through SFT and RL before consolidation into a single model via **on-policy distillation** [2603.27703]. The report formalizes RL training samples as
$$
\mathcal{D}_{\text{RL}} = \left\{ \langle \mathcal{E}, \mathcal{T}_{\text{tools}}, \mathcal{S}_{\text{agent}}, \mathcal{I}_{\text{task}}, \mathcal{V}_{\text{verifier}} \rangle \right\},
$$
emphasizing that an agentic example is not merely a prompt-response pair but an environment, toolset, scaffold, task, and verifier.

KAT-Coder-V2 also introduces two infrastructure and optimization components. **KwaiEnv** is a modular environment layer that supports datasets, verifiers, scaffolds, sandboxes, and trajectory management, with the report stating support for **tens of thousands of concurrent sandbox instances**. **MCLA**, or **Monte-Carlo Log-probability Averaging**, addresses **MoE RL instability** by averaging repeated log-probability estimates,
$$
\bar{\log \pi}(a) = \frac{1}{K} \sum_{k=1}^{K} \log \pi^{(k)}(a), \quad K=8,
$$
while **Tree Training** removes redundant computation on tree-structured agent trajectories and is reported to yield **up to 6.2× speedup** [2603.27703].

## 6. Empirical profile, deployment, limitations, and disambiguation

KAT-V1-40B is reported as highly competitive for its scale on reasoning-heavy tasks. The main benchmark table lists **77.8** on **MMLU-Pro**, **91.2** on **DROP**, **75.1** on **GPQA-Diamond**, **97.4** on **MATH500**, **93.3** on **AIME2024**, **88.1** on **AIME2025**, **95.1** on **HumanEval**, and **74.2** on **LiveCodeBench** [2507.08297]. The paper’s strongest comparative claims are on difficult math and coding tasks, especially against larger baselines such as **DeepSeek-R1-0528** and **Qwen3-235B-A22B**. Token efficiency is a second major theme: the report claims token reduction by **up to approximately 30%**, cites **27.3% fewer tokens** in a specific AIME comparison, and elsewhere reports average savings ranging from **11.6% to 89.9%** depending on setting [2507.08297].

The coding branch is evaluated on explicitly agentic benchmarks. KAT-Coder is reported at **73.4** on **SWE-Bench Verified**, **86.0** on **IFEval**, **62.3** on **TAU2-Bench Retail**, **72.5** on **AIME 2025**, **48.2** on **LiveCodeBench V6**, **96.3** on **HumanEval**, and **68.2** on **GPQA-Diamond** [2510.18779]. KAT-Coder-V2 further reports **79.6%** on **SWE-bench Verified**, **88.7** on **PinchBench**, leading results across all three reported **frontend aesthetics** scenarios, **46.8** on **Terminal-Bench Hard**, and **93.9** on **$\tau^2$-Bench** [2603.27703]. These reports collectively position KAT as a family that moves from adaptive reasoning to full agentic software work.

The deployment evidence is partly benchmark-based and partly anecdotal. KAT-V1 is described as already deployed in **Kwaipilot**, where the paper uses examples of simple queries answered in think-off mode and complex SQL tuning tasks answered in think-on mode. KAT-Coder’s final adaptation stage is explicitly aligned to **production-grade IDE environments**, and KAT-Dev is the open release meant to operationalize that deployment line [2507.08297] [2510.18779].

The reports are also marked by several disclosure asymmetries. KAT-V1 does not provide explicit equations for the distillation loss, the exact judge/answer reward composition, or the full Step-SRPO optimization. KAT-Coder leaves unspecified the base architecture family, tokenizer, context window, optimizer, batch size, hardware, and learning-rate schedule. KAT-Coder-V2 discusses MoE stabilization and large-scale RL infrastructure, but does not disclose the full MoE parameterization, routing policy, or a complete OPD objective [2507.08297] [2510.18779] [2603.27703]. This suggests that the KAT line is unusually detailed about data curation and training philosophy, but selectively detailed about some implementation internals.

Finally, **Kwaipilot-AutoThink (KAT)** should be distinguished from several unrelated arXiv systems that share the same acronym. **“KAT: A Knowledge Augmented Transformer for Vision-and-Language”** defines KAT as **Knowledge Augmented Transformer**, a multimodal OK-VQA model; the overlap is described as almost certainly **coincidental** [2112.08614]. **“KAT: Dependency-aware Automated API Testing with Large Language Models”** defines KAT as **Katalon API Testing**, an LLM-based REST API testing framework rather than an AutoThink model family [2407.10227]. The acronym therefore refers to multiple unrelated research objects across the literature; within the Kuaishou line, however, KAT consistently denotes **Kwaipilot-AutoThink** and its coding descendants.

Source: https://www.emergentmind.com/topics/kwaipilot-autothink-kat