Spontaneous Functional Differentiation in Large Language Models: A Brain-Like Intelligence Economy

The evolution of intelligence in artificial systems provides a unique opportunity to identify universal computational principles. Here we show that large language models spontaneously develop synergistic cores where information integration exceeds individual parts remarkably similar to the human brain. Using Integrated Information Decomposition across multiple architectures we find that middle layers exhibit synergistic processing while early and late layers rely on redundancy. This organization is dynamic and emerges as a physical phase transition as task difficulty increases. Crucially ablating synergistic components causes catastrophic performance loss confirming their role as the physical entity of abstract reasoning and bridging artificial and biological intelligence.

Grokking From Abstraction to Intelligence

Grokking in modular arithmetic has established itself as the quintessential fruit fly experiment, serving as a critical domain for investigating the mechanistic origins of model generalization. Despite its significance, existing research remains narrowly focused on specific local circuits or optimization tuning, largely overlooking the global structural evolution that fundamentally drives this phenomenon. We propose that grokking originates from a spontaneous simplification of internal model structures governed by the principle of parsimony. We integrate causal, spectral, and algorithmic complexity measures alongside Singular Learning Theory to reveal that the transition from memorization to generalization corresponds to the physical collapse of redundant manifolds and deep information compression, offering a novel perspective for understanding the mechanisms of model overfitting and generalization.

KAT-Coder-V2 Technical Report

We present KAT-Coder-V2, an agentic coding model developed by the KwaiKAT team at Kuaishou. KAT-Coder-V2 adopts a "Specialize-then-Unify" paradigm that decomposes agentic coding into five expert domains - SWE, WebCoding, Terminal, WebSearch, and General - each undergoing independent supervised fine-tuning and reinforcement learning, before being consolidated into a single model via on-policy distillation. We develop KwaiEnv, a modular infrastructure sustaining tens of thousands of concurrent sandbox instances, and scale RL training along task complexity, intent alignment, and scaffold generalization. We further propose MCLA for stabilizing MoE RL training and Tree Training for eliminating redundant computation over tree-structured trajectories with up to 6.2x speedup. KAT-Coder-V2 achieves 79.6% on SWE-bench Verified (vs. Claude Opus 4.6 at 80.8%), 88.7 on PinchBench (surpassing GLM-5 and MiniMax M2.7), ranks first across all three frontend aesthetics scenarios, and maintains strong generalist scores on Terminal-Bench Hard (46.8) and tau^2-Bench (93.9). Our model is publicly available at https://streamlake.com/product/kat-coder.

TSC: Topology-Conditioned Stackelberg Coordination for Multi-Agent Reinforcement Learning in Interactive Driving

Safe and efficient autonomous driving in dense traffic is fundamentally a decentralized multi-agent coordination problem, where interactions at conflict points such as merging and weaving must be resolved reliably under partial observability. With only local and incomplete cues, interaction patterns can change rapidly, often causing unstable behaviors such as oscillatory yielding or unsafe commitments. Existing multi-agent reinforcement learning (MARL) approaches either adopt synchronous decision-making, which exacerbate non-stationarity, or depend on centralized sequencing mechanisms that scale poorly as traffic density increases. To address these limitations, we propose Topology-conditioned Stackelberg Coordination (TSC), a learning framework for decentralized interactive driving under communication-free execution, which extracts a time-varying directed priority graph from braid-inspired weaving relations between trajectories, thereby defining local leader-follower dependencies without constructing a global order of play. Conditioned on this graph, TSC endogenously factorizes dense interactions into graph-local Stackelberg subgames and, under centralized training and decentralized execution (CTDE), learns a sequential coordination policy that anticipates leaders via action prediction and trains followers through action-conditioned value learning to approximate local best responses, improving training stability and safety in dense traffic. Experiments across four dense traffic scenarios show that TSC achieves superior performance over representative MARL baselines across key metrics, most notably reducing collisions while maintaining competitive traffic efficiency and control smoothness.

MalRAG: A Retrieval-Augmented LLM Framework for Open-set Malicious Traffic Identification

Fine-grained identification of IDS-flagged suspicious traffic is crucial in cybersecurity. In practice, cyber threats evolve continuously, making the discovery of novel malicious traffic a critical necessity as well as the identification of known classes. Recent studies have advanced this goal with deep models, but they often rely on task-specific architectures that limit transferability and require per-dataset tuning. In this paper we introduce MalRAG, the first LLM driven retrieval-augmented framework for open-set malicious traffic identification. MalRAG freezes the LLM and operates via comprehensive traffic knowledge construction, adaptive retrieval, and prompt engineering. Concretely, we construct a multi-view traffic database by mining prior malicious traffic from content, structural, and temporal perspectives. Furthermore, we introduce a Coverage-Enhanced Retrieval Algorithm that queries across these views to assemble the most probable candidates, thereby improving the inclusion of correct evidence. We then employ Traffic-Aware Adaptive Pruning to select a variable subset of these candidates based on traffic-aware similarity scores, suppressing incorrect matches and yielding reliable retrieved evidence. Moreover, we develop a suite of guidance prompts where task instruction, evidence referencing, and decision guidance are integrated with the retrieved evidence to improve LLM performance. Across diverse real-world datasets and settings, MalRAG delivers state-of-the-art results in both fine-grained identification of known classes and novel malicious traffic discovery. Ablation and deep-dive analyses further show that MalRAG effective leverages LLM capabilities yet achieves open-set malicious traffic identification without relying on a specific LLM.

T2VParser: Adaptive Decomposition Tokens for Partial Alignment in Text to Video Retrieval

Text-to-video retrieval essentially aims to train models to align visual content with textual descriptions accurately. Due to the impressive general multimodal knowledge demonstrated by image-text pretrained models such as CLIP, existing work has primarily focused on extending CLIP knowledge for video-text tasks. However, videos typically contain richer information than images. In current video-text datasets, textual descriptions can only reflect a portion of the video content, leading to partial misalignment in video-text matching. Therefore, directly aligning text representations with video representations can result in incorrect supervision, ignoring the inequivalence of information. In this work, we propose T2VParser to extract multiview semantic representations from text and video, achieving adaptive semantic alignment rather than aligning the entire representation. To extract corresponding representations from different modalities, we introduce Adaptive Decomposition Tokens, which consist of a set of learnable tokens shared across modalities. The goal of T2VParser is to emphasize precise alignment between text and video while retaining the knowledge of pretrained models. Experimental results demonstrate that T2VParser achieves accurate partial alignment through effective cross-modal content decomposition. The code is available at https://github.com/Lilidamowang/T2VParser.

Respond to Change with Constancy: Instruction-tuning with LLM for Non-I.I.D. Network Traffic Classification

Encrypted traffic classification is highly challenging in network security due to the need for extracting robust features from content-agnostic traffic data. Existing approaches face critical issues: (i) Distribution drift, caused by reliance on the closedworld assumption, limits adaptability to realworld, shifting patterns; (ii) Dependence on labeled data restricts applicability where such data is scarce or unavailable. Large language models (LLMs) have demonstrated remarkable potential in offering generalizable solutions across a wide range of tasks, achieving notable success in various specialized fields. However, their effectiveness in traffic analysis remains constrained by challenges in adapting to the unique requirements of the traffic domain. In this paper, we introduce a novel traffic representation model named Encrypted Traffic Out-of-Distribution Instruction Tuning with LLM (ETooL), which integrates LLMs with knowledge of traffic structures through a self-supervised instruction tuning paradigm. This framework establishes connections between textual information and traffic interactions. ETooL demonstrates more robust classification performance and superior generalization in both supervised and zero-shot traffic classification tasks. Notably, it achieves significant improvements in F1 scores: APP53 (I.I.D.) to 93.19%(6.62%) and 92.11%(4.19%), APP53 (O.O.D.) to 74.88%(18.17%) and 72.13%(15.15%), and ISCX-Botnet (O.O.D.) to 95.03%(9.16%) and 81.95%(12.08%). Additionally, we construct NETD, a traffic dataset designed to support dynamic distributional shifts, and use it to validate ETooL's effectiveness under varying distributional conditions. Furthermore, we evaluate the efficiency gains achieved through ETooL's instruction tuning approach.

Feasibility-Aware Pessimistic Estimation: Toward Long-Horizon Safety in Offline RL

Offline safe reinforcement learning(OSRL) derives constraint-satisfying policies from pre-collected datasets, offers a promising avenue for deploying RL in safety-critical real-world domains such as robotics. However, the majority of existing approaches emphasize only short-term safety, neglecting long-horizon considerations. Consequently, they may violate safety constraints and fail to ensure sustained protection during online deployment. Moreover, the learned policies often struggle to handle states and actions that are not present or out-of-distribution(OOD) from the offline dataset, and exhibit limited sample efficiency. To address these challenges, we propose a novel framework Feasibility-Aware offline Safe Reinforcement Learning with CVAE-based Pessimism (FASP). First, we employ Hamilton-Jacobi (H-J) reachability analysis to generate reliable safety labels, which serve as supervisory signals for training both a conditional variational autoencoder (CVAE) and a safety classifier. This approach not only ensures high sampling efficiency but also provides rigorous long-horizon safety guarantees. Furthermore, we utilize pessimistic estimation methods to estimate the Q-value of reward and cost, which mitigates the extrapolation errors induces by OOD actions, and penalize unsafe actions to enabled the agent to proactively avoid high-risk behaviors. Moreover, we theoretically prove the validity of this pessimistic estimation. Extensive experiments on DSRL benchmarks demonstrate that FASP algorithm achieves competitive performance across multiple experimental tasks, particularly outperforming state-of-the-art algorithms in terms of safety.

Stop Summation: Min-Form Credit Assignment Is All Process Reward Model Needs for Reasoning

Process reward models (PRMs) have proven effective for test-time scaling of Large Language Models (LLMs) on challenging reasoning tasks. However, reward hacking issues with PRMs limit their successful application in reinforcement fine-tuning. In this paper, we identify the main cause of PRM-induced reward hacking: the canonical summation-form credit assignment in reinforcement learning (RL), which defines the value as cumulative gamma-decayed future rewards, easily induces LLMs to hack steps with high rewards. To address this, we propose PURE: Process sUpervised Reinforcement lEarning. The key innovation of PURE is a min-form credit assignment that formulates the value function as the minimum of future rewards. This method significantly alleviates reward hacking by limiting the value function range and distributing advantages more reasonably. Through extensive experiments on 3 base models, we show that PRM-based approaches enabling min-form credit assignment achieve comparable reasoning performance to verifiable reward-based methods within only 30% steps. In contrast, the canonical sum-form credit assignment collapses training even at the beginning! Additionally, when we supplement PRM-based fine-tuning with just 10% verifiable rewards, we further alleviate reward hacking and produce the best fine-tuned model based on Qwen2.5-Math-7B in our experiments, achieving 82.5% accuracy on AMC23 and 53.3% average accuracy across 5 benchmarks. Moreover, we summarize the observed reward hacking cases and analyze the causes of training collapse. Code and models are available at https://github.com/CJReinforce/PURE.

Missing Target-Relevant Information Prediction with World Model for Accurate Zero-Shot Composed Image Retrieval

Zero-Shot Composed Image Retrieval (ZS-CIR) involves diverse tasks with a broad range of visual content manipulation intent across domain, scene, object, and attribute. The key challenge for ZS-CIR tasks is to modify a reference image according to manipulation text to accurately retrieve a target image, especially when the reference image is missing essential target content. In this paper, we propose a novel prediction-based mapping network, named PrediCIR, to adaptively predict the missing target visual content in reference images in the latent space before mapping for accurate ZS-CIR. Specifically, a world view generation module first constructs a source view by omitting certain visual content of a target view, coupled with an action that includes the manipulation intent derived from existing image-caption pairs. Then, a target content prediction module trains a world model as a predictor to adaptively predict the missing visual information guided by user intention in manipulating text at the latent space. The two modules map an image with the predicted relevant information to a pseudo-word token without extra supervision. Our model shows strong generalization ability on six ZS-CIR tasks. It obtains consistent and significant performance boosts ranging from 1.73% to 4.45% over the best methods and achieves new state-of-the-art results on ZS-CIR. Our code is available at https://github.com/Pter61/predicir.