Complex reinforcement learning environments frequently employ multi-task and mixed-reward formulations. In these settings, heterogeneous reward distributions and correlated reward dimensions often destabilize the construction of scalar advantages. To address these challenges, we propose Reward-Decorrelated Policy Optimization (RDPO), a reward-processing method designed to explicitly target both failure modes. RDPO first utilizes Magnitude-Aware Quantile normalization to stabilize prompt-level advantage allocation across binary, fractional, and continuous rewards. It then applies Mahalanobis whitening within each active reward subspace to mitigate correlation redundancy prior to aggregation. When applied during the post-training of LongCat-Flash, RDPO enhances instruction following, writing quality, and robustness to hard prompts while remaining broadly competitive on reasoning and coding evaluations.

As reinforcement learning continues to scale the training of large language model-based agents, reliably verifying agent behaviors in complex environments has become increasingly challenging. Existing approaches rely on rule-based verifiers or LLM-as-a-Judge models, which struggle to generalize beyond narrow domains. Agent-as-a-Judge addresses this limitation by actively interacting with environments and tools to acquire verifiable evidence, yet its capabilities remain underexplored. We introduce a benchmark AJ-Bench to systematically evaluate Agent-as-a-Judge across three domains-search, data systems, and graphical user interfaces-comprising 155 tasks and 516 annotated trajectories. The benchmark comprehensively assesses judge agents' abilities in information acquisition, state verification, and process verification. Experiments demonstrate consistent performance gains over LLM-as-a-Judge baselines, while also revealing substantial open challenges in agent-based verification. Our data and code are available at https://aj-bench.github.io/.

Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek-R1) have led to remarkable improvements through long Chain-of-Thought (CoT). However, existing benchmarks mainly focus on immediate, single-horizon tasks, failing to adequately evaluate models’ ability to understand and respond to complex, long-horizon scenarios. To address this incomplete evaluation of Large Reasoning Models (LRMs), we propose R-HORIZON, a method designed to stimulate long-horizon reasoning behaviors in LRMs through query composition. Based on R-HORIZON, we construct a long-horizon reasoning benchmark, comprising complex multi-step reasoning tasks with interdependent problems that span long reasoning horizons. Through comprehensive evaluation of LRMs using the R-HORIZON benchmark, we find that even the most advanced LRMs suffer significant performance degradation. Our analysis reveals that LRMs exhibit limited effective reasoning length and struggle to allocate thinking budget across multiple problems appropriately. Recognizing these limitations, we use R-HORIZON to construct long-horizon reasoning data for reinforcement learning with verified rewards (RLVR). Compared to training with single-horizon data, RLVR with R-HORIZON not only substantially improves performance on the multi-horizon reasoning tasks, but also promotes accuracy on standard reasoning tasks (+7.5 on AIME2024). These results position R-HORIZON as a scalable, controllable, and low-cost paradigm for enhancing and evaluating the long-horizon reasoning capabilities of LRMs.

A persistent skill library allows language model agents to reuse successful strategies across tasks. Maintaining such a library requires three coupled capabilities. The agent selects a relevant skill, utilizes it during execution, and distills new skills from experience. Existing methods optimize these capabilities in isolation or with separate reward sources, resulting in partial and conflicting evolution. We propose Skill1, a framework that trains a single policy to co-evolve skill selection, utilization, and distillation toward a shared task-outcome objective. The policy generates a query to search the skill library, re-ranks candidates to select one, solves the task conditioned on it, and distills a new skill from the trajectory. All learning derives from a single task-outcome signal. Its low-frequency trend credits selection and its high-frequency variation credits distillation. Experiments on ALFWorld and WebShop show that Skill1 outperforms prior skill-based and reinforcement learning baselines. Training dynamics confirm the co-evolution of the three capabilities, and ablations show that removing any credit signal degrades the evolution.

As the parameter size of large language models (LLMs) continues to grow, the latency of autoregressive inference increases due to memory-bound computational inefficiency. To address this, speculative decoding has been proposed, where a large target model verifies multiple tokens generated in parallel by a smaller draft model. However, the performance of speculative decoding is fundamentally limited by the draft model’s capacity, which stems from the parameter gap between the two models. To overcome this limitation, we propose RepSpec, which combines structural re-parameterization with draft model training. During training, redundant linear structures are introduced and later merged into the backbone network during inference, thus enhancing the draft model’s training effectiveness without increasing inference cost. By applying our method to improve the current state-of-the-art approach, EAGLE, we achieve a significant improvement in accepted sequence length. Furthermore, considering the specific characteristics of the speculative decoding scenario, we explore a hybrid training strategy that combines linear and nonlinear structures, which yields a further improvement in acceptance length.

The increasing complexity of agentic tasks has led to rapidly growing trajectory lengths, which poses significant challenges for large language model (LLM) based agents with fixed context windows. Existing context management techniques, such as truncation and summarization, suffer from inherent inflexibility and irreversibility: once information is discarded or compressed, it cannot be recovered even when it becomes critically relevant in later decision steps. To address these limitations, we propose the Adaptive Context Elasticizer (ACE), a plug-and-play module that elastically orchestrates historical step information into the agent's context at each decision step. ACE maintains a lossless message maintenance layer that stores both raw messages and compressed abstractions for each historical step, while a context orchestration layer adaptively assigns each step an elastic type as raw, abstract, or drop, at every decision step based on the current task state. This reversible design ensures that the main LLM always receives a compact yet information-rich context. We adapt ACE to four diverse agent frameworks, including ReAct, DeepAgent, WebThinker, and MiroFlow, without training or architectural modifications. Experiments show that ACE consistently outperforms truncation and summarization baselines, and brings consistent performance gains across all four agent frameworks.

Reinforcement learning with verifiable rewards (RLVR) broadcasts a single response-level reward to every token, while on-policy distillation (OPD) scores each token against a stronger teacher for a dense advantage but caps performance at teacher quality and discourages exploration beyond it. Their complementarity makes combining RLVR and OPD promising, but we find that fusing the two advantages with a fixed coefficient triggers entropy collapse from two miscalibrations: a magnitude mismatch, where token-level OPD advantages can spike far beyond the bounded RLVR advantage and erase its signal, and a temporal mismatch, where sustained full-strength OPD keeps pulling the student toward the teacher and limits exploration needed to surpass it. We propose SAF, a Stable Advantage Fusion framework that resolves both issues via a lightweight, four-stage pipeline applied only to the OPD advantage: a sparsify-then-compress mechanism for magnitude control paired with a warm-up-then-anneal mechanism for temporal control, with each stage independently switchable and adding negligible overhead. Instantiating RLVR with GRPO, we evaluate SAF across seven mathematical reasoning and code generation benchmarks with Qwen3-1.7B/4B/8B: SAF avoids entropy collapse and consistently outperforms fixed-coefficient GRPO+OPD fusion, improving the aggregate score by 0.51-2.70

Existing Reinforcement Learning with Verifiable Rewards (RLVR) algorithms, such as GRPO, rely on rigid, uniform, and symmetric trust region mechanisms that are fundamentally misaligned with the complex optimization dynamics of Large Language Models (LLMs). In this paper, we identify three critical challenges in these methods: (1) inefficient gradient utilization caused by the binary cutoff of hard clipping, (2) insensitive probability mass arising from uniform ratio constraints that ignore the token distribution, and (3) asymmetric signal reliability stemming from the disparate credit assignment ambiguity between positive and negative samples. To bridge these gaps, we propose Mass-Adaptive Soft Policy Optimization (MASPO), a unified framework designed to harmonize these three dimensions. MASPO integrates a differentiable soft Gaussian gating to maximize gradient utility, a mass-adaptive limiter to balance exploration across the probability spectrum, and an asymmetric risk controller to align update magnitudes with signal confidence. Extensive evaluations demonstrate that MASPO serves as a robust, all-in-one RLVR solution, significantly outperforming baselines. Our code is available at: https://github.com/FlyTune/MASPO-RL.

On-policy distillation (OPD) trains a student on its own trajectories with token-level teacher feedback and often outperforms off-policy distillation and standard reinforcement learning. However, we find that its standard advantage weighted policy gradient suffers from three structural weaknesses, including high variance updates, vanishing gradients in zero-advantage regions, and exploration bottlenecks when corrective signals are insufficient.We therefore propose Asymmetric On-Policy Distillation (AOPD), which replaces ineffective negative reinforcement with localized divergence minimization in non-positive advantage regions while preserving positive reinforcement learning. Experiments on mathematical reasoning benchmarks show that AOPD consistently outperforms standard OPD, with average gains of 4.09 / 8.34 under strong / weak initialization, respectively. AOPD also maintains higher policy entropy during training and better capability retention during sequential tool-use adaptation.

Long-horizon agentic reasoning necessitates effectively compressing growing interaction histories into a limited context window. Most existing memory systems serialize history as text, where token-level cost is uniform and scales linearly with length, often spending scarce budget on low-value details. To this end, we introduce , a multimodal memory agent that improves long-horizon reasoning under tight context budgets by allocating memory space with adaptive information density through visual layout. Concretely, MemOCR maintains a structured rich-text memory (e.g., headings, highlights) and renders it into an image that the agent consults for memory access, visually prioritizing crucial evidence while aggressively compressing auxiliary details. To ensure robustness across varying memory budgets, we train MemOCR with reinforcement learning under budget-aware objectives that expose the agent to diverse compression levels. Across long-context multi-hop and single-hop question-answering benchmarks, MemOCR outperforms strong text-based baselines and achieves more effective context utilization under extreme budgets.

Enabling Large Language Models (LLMs) to effectively utilize tools in multi-turn interactions is essential for building capable autonomous agents. However, acquiring diverse and realistic multi-turn tool-use data remains a significant challenge. In this work, we propose a novel text-based paradigm. We observe that textual corpora naturally contain rich, multi-step problem-solving experiences, which can serve as an untapped, scalable, and authentic data source for multi-turn tool-use tasks. Based on this insight, we introduce GEM, a data synthesis pipeline that enables the generation and extraction of multi-turn tool-use trajectories from text corpora through a four-stage process: relevance filtering, workflow tool extraction, trajectory grounding, and complexity refinement. To reduce the computational cost, we further train a specialized Trajectory Synthesizer via supervised fine-tuning. This model distills the complex generation pipeline into an efficient, end-to-end trajectory generator. Experiments demonstrate that our GEM-32B achieve a 14.9% improvement on the BFCL V3 Multi-turn benchmark. Our models partially surpass the performance of models trained on -bench (Airline and Retail) in-domain data, highlighting the superior generalization capability derived from our text-based synthesis paradigm. Notably, our Trajectory Synthesizer matches the quality of the full pipeline while significantly reducing inference latency and costs.

Autonomous agents are increasingly capable of improving models, systems, and other technical artifacts through long-horizon experimentation. To understand the current state of this capability, however, evaluation must go beyond final scores, which neither reveal where progress is gained or lost nor indicate whether accumulated experience improves later decisions. We therefore present a systematic evaluation of seven frontier models on 36 long-horizon tasks based on a new framework that uses rule-based metrics to characterize within-run behavior through Solution Framing, Execution, and Feedback Control and controlled comparisons to assess experience reuse within and across tasks. The results show that current agents operate more like engineering optimizers than fully autonomous researchers: they can formulate and implement practical solutions, but their performance varies substantially across runs, their strongest solutions mainly adapt or combine established techniques, and genuine methodological novelty remains rare. Detailed analysis reveals that observed performance is shaped by multiple factors, including distinct process bottlenecks behind similar final outcomes, experience reuse that can help or mislead subsequent decisions, and harness designs that affect performance stability. These findings suggest concrete directions for improving model training, inference-time strategies, experience management, and harness design.

Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for Large Language Model (LLM) reasoning, yet current methods face key challenges in resource allocation and policy optimization dynamics: (i) uniform rollout allocation ignores gradient variance heterogeneity across problems, and (ii) the softmax policy structure causes gradient attenuation for high-confidence correct actions, while excessive gradient updates may destabilize training. Therefore, we propose DynaMO, a theoretically-grounded dual-pronged optimization framework. At the sequence level, we prove that uniform allocation is suboptimal and derive variance-minimizing allocation from the first principle, establishing Bernoulli variance as a computable proxy for gradient informativeness. At the token level, we develop gradient-aware advantage modulation grounded in theoretical analysis of gradient magnitude bounds. Our framework compensates for gradient attenuation of high-confidence correct actions while utilizing entropy changes as computable indicators to stabilize excessive update magnitudes. Extensive experiments conducted on a diverse range of mathematical reasoning benchmarks demonstrate consistent improvements over strong RLVR baselines. Our implementation is available at: \href{https://anonymous.4open.science/r/dynamo-680E/README.md}{https://anonymous.4open.science/r/dynamo}.

Softmax attention stores key and value vectors for every preceding token, causing inference memory to grow with sequence length. Recent language models incorporating Gated DeltaNet (GDN) or Kimi Delta Attention (KDA) reduce this cost by replacing the KV cache in most layers with fixed-size recurrent states. However, these recurrent states are commonly stored in FP32 and consume substantial GPU memory; their updates are memory-bandwidth bound and contribute significantly to decoding latency. To our knowledge, we are the first to study post-training quantization of recurrent states in GDN and KDA based language models. We find that uniform quantization provides a poor accuracy–storage trade-off: INT8 and FP8 already degrade accuracy on complex reasoning tasks, while INT4 and NVFP4 reduce it to near zero. We further find that most quantization-error energy is concentrated in a small subset of channels and that the relative decay strength of state channels remains stable across prompts and tasks. Motivated by these findings, DAMP uses both quantization-error energy and decay-based persistence to identify high-risk channels during offline calibration. It stores these channels at higher precision and the remainder in INT8. We evaluate DAMP on Qwen3.6-35B and Kimi-Linear-48B across six benchmarks covering mathematical reasoning, general reasoning, and code generation. At 9.9 bits per state value, DAMP maintains average accuracy close to the FP32 baseline. DAMP reduces recurrent-state storage by 69.1

While sparse attention mitigates the computational bottleneck of long-context LLM training, its distributed training process exhibits extreme heterogeneity in both 1) sequence length and 2) sparsity sensitivity, leading to a severe imbalance problem and sub-optimal model accuracy. Existing algorithms and training frameworks typically focus on single issue, failing to systematically co-optimize these two problems. Therefore, we propose SparseBalance, a novel algorithm-system co-design framework, which exploits the sparsity and sequence heterogeneity to optimize model accuracy and system efficiency jointly. First, we propose workload-aware dynamic sparsity tuning, which employs a bidirectional sparsity adjustment to eliminate stragglers and exploit inherent bubbles for free accuracy. Second, we propose a sparsity-aware batching strategy to achieve coarse-grained balance, which complements dynamic sparsity tuning. Experimental results demonstrate that SparseBalance achieves up to a 1.33× end-to-end speedup while still improving the long-context capability by 0.46% on the LongBench benchmark.

Attention–Feed-Forward Network (FFN) Disaggregation (AFD) is emerging as a promising architecture for serving Mixture-of-Experts (MoE) language models. While existing AFD systems improve the efficiency of disaggregated execution, they leave a deployment question unanswered: under the same model, workload, time-per-output-token (TPOT) service-level objective (SLO), hardware budget, hardware catalog, and runtime capabilities, does AFD provide higher throughput than the best collocated deployment? Answering this question requires jointly optimizing hardware assignment and deployment organization for both architectures, making exhaustive provisioning prohibitively expensive. We present AFD-Ledger, an offline analytical provisioning system that independently provisions AFD and collocated deployments using an analytical execution model and an evaluation-bounded hardware search. Across deployment spaces where exhaustive provisioning is feasible, AFD-Ledger reduces complete deployment evaluations by 68.8

Recent advances in large language models have enabled LLM-based agents to achieve strong performance on a variety of benchmarks. However, their performance in real-world deployments often that observed on benchmark settings, especially in complex and imperfect environments. This discrepancy largely arises because prevailing training and evaluation paradigms are typically built on idealized assumptions, overlooking the inherent stochasticity and noise present in real-world interactions. To bridge this gap, we introduce AgentNoiseBench, a framework for systematically evaluating the robustness of agentic models under noisy environments. We first conduct an in-depth analysis of biases and uncertainties in real-world scenarios and categorize environmental noise into two primary types: user-noise and tool-noise. Building on this analysis, we develop an automated pipeline that injects controllable noise into existing agent-centric benchmarks while preserving task solvability. Leveraging this pipeline, we perform extensive evaluations across a wide range of models with diverse architectures and parameter scales. Our results reveal consistent performance variations under different noise conditions, highlighting the sensitivity of current agentic models to realistic environmental perturbations.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key approach for enhancing LLM reasoning.However, standard frameworks like Group Relative Policy Optimization (GRPO) typically employ a uniform rollout budget, leading to resource inefficiency. Moreover, existing adaptive methods often rely on instance-level metrics, such as task pass rates, failing to capture the model's dynamic learning state. To address these limitations, we propose CoBA-RL, a reinforcement learning algorithm designed to adaptively allocate rollout budgets based on the model's evolving capability. Specifically, CoBA-RL utilizes a Capability-Oriented Value function to map tasks to their potential training gains and employs a heap-based greedy strategy to efficiently self-calibrate the distribution of computational resources to samples with high training value. Extensive experiments demonstrate that our approach effectively orchestrates the trade-off between exploration and exploitation, delivering consistent generalization improvements across multiple challenging benchmarks. These findings underscore that quantifying sample training value and optimizing budget allocation are pivotal for advancing LLM post-training efficiency.

End-to-end Spoken Language Models (SLMs) hold great potential for paralinguistic perception, and numerous studies have aimed to enhance their capabilities, particularly for empathetic dialogue. However, current approaches largely depend on rigid supervised signals, such as ground-truth response in supervised fine-tuning or preference scores in reinforcement learning. Such reliance is fundamentally limited for modeling complex empathy, as there is no single "correct" response and a simple numerical score cannot fully capture the nuances of emotional expression or the appropriateness of empathetic behavior. To address these limitations, we sequentially introduce EmpathyEval, a descriptive natural-language-based evaluation model for assessing empathetic quality in spoken dialogues. Building upon EmpathyEval, we propose ReEmpathy, an end-to-end SLM that enhances empathetic dialogue through a novel Empathetic Self-Reflective Alternating Inference mechanism, which interleaves spoken response generation with free-form, empathy-related reflective reasoning. Extensive experiments demonstrate that ReEmpathy substantially improves empathy-sensitive spoken dialogue by enabling reflective reasoning, offering a promising approach toward more emotionally intelligent and empathy-aware human-computer interactions.