Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis

Deploying Large Language Models to data-scarce programming domains poses significant challenges, particularly for kernel synthesis on emerging Domain-Specific Architectures where a "Data Wall" limits available training data. While models excel on data-rich platforms like CUDA, they suffer catastrophic performance drops on data-scarce ecosystems such as NPU programming. To overcome this cold-start barrier without expensive fine-tuning, we introduce EvoKernel, a self-evolving agentic framework that automates the lifecycle of kernel synthesis from initial drafting to continual refining. EvoKernel addresses this by formulating the synthesis process as a memory-based reinforcement learning task. Through a novel value-driven retrieval mechanism, it learns stage-specific Q-values that prioritize experiences based on their contribution to the current objective, whether bootstrapping a feasible draft or iteratively refining latency. Furthermore, by enabling cross-task memory sharing, the agent generalizes insights from simple to complex operators. By building an NPU variant of KernelBench and evaluating on it, EvoKernel improves frontier models' correctness from 11.0% to 83.0% and achieves a median speedup of 3.60x over initial drafts through iterative refinement. This demonstrates that value-guided experience accumulation allows general-purpose models to master the kernel synthesis task on niche hardware ecosystems. Our official page is available at https://evokernel.zhuo.li.

Revealing Behavioral Plasticity in Large Language Models: A Token-Conditional Perspective

In this work, we reveal that Large Language Models (LLMs) possess intrinsic behavioral plasticity-akin to chameleons adapting their coloration to environmental cues-that can be exposed through token-conditional generation and stabilized via reinforcement learning. Specifically, by conditioning generation on carefully selected token prefixes sampled from responses exhibiting desired behaviors, LLMs seamlessly adapt their behavioral modes at inference time (e.g., switching from step-by-step reasoning to direct answering) without retraining. Based on this insight, we propose Token-Conditioned Reinforcement Learning (ToCoRL), a principled framework that leverages RL to internalize this chameleon-like plasticity, transforming transient inference-time adaptations into stable and learnable behavioral patterns. ToCoRL guides exploration with token-conditional generation and keep enhancing exploitation, enabling emergence of appropriate behaviors. Extensive experiments show that ToCoRL enables precise behavioral control without capability degradation. Notably, we show that large reasoning models, while performing strongly on complex mathematics, can be effectively adapted to excel at factual question answering, which was a capability previously hindered by their step-by-step reasoning patterns.

PhotoBench: Beyond Visual Matching Towards Personalized Intent-Driven Photo Retrieval

Personal photo albums are not merely collections of static images but living, ecological archives defined by temporal continuity, social entanglement, and rich metadata, which makes the personalized photo retrieval non-trivial. However, existing retrieval benchmarks rely heavily on context-isolated web snapshots, failing to capture the multi-source reasoning required to resolve authentic, intent-driven user queries. To bridge this gap, we introduce PhotoBench, the first benchmark constructed from authentic, personal albums. It is designed to shift the paradigm from visual matching to personalized multi-source intent-driven reasoning. Based on a rigorous multi-source profiling framework, which integrates visual semantics, spatial-temporal metadata, social identity, and temporal events for each image, we synthesize complex intent-driven queries rooted in users' life trajectories. Extensive evaluation on PhotoBench exposes two critical limitations: the modality gap, where unified embedding models collapse on non-visual constraints, and the source fusion paradox, where agentic systems perform poor tool orchestration. These findings indicate that the next frontier in personal multimodal retrieval lies beyond unified embeddings, necessitating robust agentic reasoning systems capable of precise constraint satisfaction and multi-source fusion. Our PhotoBench is available.

MuonRec: Shifting the Optimizer Paradigm Beyond Adam in Scalable Generative Recommendation

Recommender systems (RecSys) are increasingly emphasizing scaling, leveraging larger architectures and more interaction data to improve personalization. Yet, despite the optimizer's pivotal role in training, modern RecSys pipelines almost universally default to Adam/AdamW, with limited scrutiny of whether these choices are truly optimal for recommendation. In this work, we revisit optimizer design for scalable recommendation and introduce MuonRec, the first framework that brings the recently proposed Muon optimizer to RecSys training. Muon performs orthogonalized momentum updates for 2D weight matrices via Newton-Schulz iteration, promoting diverse update directions and improving optimization efficiency. We develop an open-source training recipe for recommendation models and evaluate it across both traditional sequential recommenders and modern generative recommenders. Extensive experiments demonstrate that MuonRec reduces converged training steps by an average of 32.4\% while simultaneously improving final ranking quality. Specifically, MuonRec yields consistent relative gains in NDCG@10, averaging 12.6\% across all settings, with particularly pronounced improvements in generative recommendation models. These results consistently outperform strong Adam/AdamW baselines, positioning Muon as a promising new optimizer standard for RecSys training. Our code is available.

Beyond Imitation: Reinforcement Learning-Based Sim-Real Co-Training for VLA Models

Simulation offers a scalable and low-cost way to enrich vision-language-action (VLA) training, reducing reliance on expensive real-robot demonstrations. However, most sim-real co-training methods rely on supervised fine-tuning (SFT), which treats simulation as a static source of demonstrations and does not exploit large-scale closed-loop interaction. Consequently, real-world gains and generalization are often limited. In this paper, we propose an \underline{\textit{RL}}-based sim-real \underline{\textit{Co}}-training \modify{(RL-Co)} framework that leverages interactive simulation while preserving real-world capabilities. Our method follows a generic two-stage design: we first warm-start the policy with SFT on a mixture of real and simulated demonstrations, then fine-tune it with reinforcement learning in simulation while adding an auxiliary supervised loss on real-world data to anchor the policy and mitigate catastrophic forgetting. We evaluate our framework on four real-world tabletop manipulation tasks using two representative VLA architectures, OpenVLA and $π_{0.5}$, and observe consistent improvements over real-only fine-tuning and SFT-based co-training, including +24% real-world success on OpenVLA and +20% on $π_{0.5}$. Beyond higher success rates, RL co-training yields stronger generalization to unseen task variations and substantially improved real-world data efficiency, providing a practical and scalable pathway for leveraging simulation to enhance real-robot deployment.

LogitsCoder: Towards Efficient Chain-of-Thought Path Search via Logits Preference Decoding for Code Generation

Code generation remains a challenging task that requires precise and structured reasoning. Existing Test Time Scaling (TTS) methods, including structured tree search, have made progress in exploring reasoning paths but still face two major challenges: (1) underthinking, where reasoning chains tend to be shallow and fail to capture the full complexity of problems; and (2) overthinking, where overly verbose reasoning leads to inefficiency and increased computational costs. To address these issues, we propose LogitsCoder, a novel framework that enhances chain-of-thought reasoning through lightweight, logit-level control mechanisms for code generation. LogitsCoder iteratively generates and refines reasoning steps by first steering token selection toward statistically preferred patterns via Logits Preference Decoding, then selecting and aggregating diverse reasoning paths using Logits Rank Based Path Selection and Thoughts Aggregation. This results in coherent and effective reasoning chains that balance depth and efficiency. Extensive experiments demonstrate that LogitsCoder produces more efficient and higher-quality reasoning chains, leading to superior code generation performance compared to baseline methods.

Plan-MCTS: Plan Exploration for Action Exploitation in Web Navigation

Large Language Models (LLMs) have empowered autonomous agents to handle complex web navigation tasks. While recent studies integrate tree search to enhance long-horizon reasoning, applying these algorithms in web navigation faces two critical challenges: sparse valid paths that lead to inefficient exploration, and a noisy context that dilutes accurate state perception. To address this, we introduce Plan-MCTS, a framework that reformulates web navigation by shifting exploration to a semantic Plan Space. By decoupling strategic planning from execution grounding, it transforms sparse action space into a Dense Plan Tree for efficient exploration, and distills noisy contexts into an Abstracted Semantic History for precise state awareness. To ensure efficiency and robustness, Plan-MCTS incorporates a Dual-Gating Reward to strictly validate both physical executability and strategic alignment and Structural Refinement for on-policy repair of failed subplans. Extensive experiments on WebArena demonstrate that Plan-MCTS achieves state-of-the-art performance, surpassing current approaches with higher task effectiveness and search efficiency.

Adaptive Milestone Reward for GUI Agents

Reinforcement Learning (RL) has emerged as a mainstream paradigm for training Mobile GUI Agents, yet it struggles with the temporal credit assignment problem inherent in long-horizon tasks. A primary challenge lies in the trade-off between reward fidelity and density: outcome reward offers high fidelity but suffers from signal sparsity, while process reward provides dense supervision but remains prone to bias and reward hacking. To resolve this conflict, we propose the Adaptive Milestone Reward (ADMIRE) mechanism. ADMIRE constructs a verifiable, adaptive reward system by anchoring trajectory to milestones, which are dynamically distilled from successful explorations. Crucially, ADMIRE integrates an asymmetric credit assignment strategy that denoises successful trajectories and scaffolds failed trajectories. Extensive experiments demonstrate that ADMIRE consistently yields over 10% absolute improvement in success rate across different base models on AndroidWorld. Moreover, the method exhibits robust generalizability, achieving strong performance across diverse RL algorithms and heterogeneous environments such as web navigation and embodied tasks.

OSCAR: Optimization-Steered Agentic Planning for Composed Image Retrieval

Composed image retrieval (CIR) requires complex reasoning over heterogeneous visual and textual constraints. Existing approaches largely fall into two paradigms: unified embedding retrieval, which suffers from single-model myopia, and heuristic agentic retrieval, which is limited by suboptimal, trial-and-error orchestration. To this end, we propose OSCAR, an optimization-steered agentic planning framework for composed image retrieval. We are the first to reformulate agentic CIR from a heuristic search process into a principled trajectory optimization problem. Instead of relying on heuristic trial-and-error exploration, OSCAR employs a novel offline-online paradigm. In the offline phase, we model CIR via atomic retrieval selection and composition as a two-stage mixed-integer programming problem, mathematically deriving optimal trajectories that maximize ground-truth coverage for training samples via rigorous boolean set operations. These trajectories are then stored in a golden library to serve as in-context demonstrations for online steering of VLM planner at online inference time. Extensive experiments on three public benchmarks and a private industrial benchmark show that OSCAR consistently outperforms SOTA baselines. Notably, it achieves superior performance using only 10% of training data, demonstrating strong generalization of planning logic rather than dataset-specific memorization.

MARTI-MARS$^2$: Scaling Multi-Agent Self-Search via Reinforcement Learning for Code Generation

While the complex reasoning capability of Large Language Models (LLMs) has attracted significant attention, single-agent systems often encounter inherent performance ceilings in complex tasks such as code generation. Multi-agent collaboration offers a promising avenue to transcend these boundaries. However, existing frameworks typically rely on prompt-based test-time interactions or multi-role configurations trained with homogeneous parameters, limiting error correction capabilities and strategic diversity. In this paper, we propose a Multi-Agent Reinforced Training and Inference Framework with Self-Search Scaling (MARTI-MARS2), which integrates policy learning with multi-agent tree search by formulating the multi-agent collaborative exploration process as a dynamic and learnable environment. By allowing agents to iteratively explore and refine within the environment, the framework facilitates evolution from parameter-sharing homogeneous multi-role training to heterogeneous multi-agent training, breaking through single-agent capability limits. We also introduce an efficient inference strategy MARTI-MARS2-T+ to fully exploit the scaling potential of multi-agent collaboration at test time. We conduct extensive experiments across varied model scales (8B, 14B, and 32B) on challenging code generation benchmarks. Utilizing two collaborating 32B models, MARTI-MARS2 achieves 77.7%, outperforming strong baselines like GPT-5.1. Furthermore, MARTI-MARS2 reveals a novel scaling law: shifting from single-agent to homogeneous multi-role and ultimately to heterogeneous multi-agent paradigms progressively yields higher RL performance ceilings, robust TTS capabilities, and greater policy diversity, suggesting that policy diversity is critical for scaling intelligence via multi-agent reinforcement learning.