Training Multi-Turn Search Agent via Contrastive Dynamic Branch Sampling

Agentic reinforcement learning has enabled large language models to perform complex multi-turn planning and tool use. However, learning in long-horizon settings remains challenging due to sparse, trajectory-level outcome rewards. While prior tree-based methods attempt to mitigate this issue, they often suffer from high variance and computational inefficiency. Through empirical analysis of search agents, We identify a common pattern: performance diverges mainly due to decisions near the tail. Motivated by this observation, we propose Branching Relative Policy Optimization (BranPO), a value-free method that provides step-level contrastive supervision without dense rewards. BranPO truncates trajectories near the tail and resamples alternative continuations to construct contrastive suffixes over shared prefixes, reducing credit ambiguity in long-horizon rollouts. To further boost efficiency and stabilize training, we introduce difficulty-aware branch sampling to adapt branching frequency across tasks, and redundant step masking to suppress uninformative actions. Extensive experiments on various question answering benchmarks demonstrate that BranPO consistently outperforms strong baselines, achieving significant accuracy gains on long-horizon tasks without increasing the overall training budget. Our code is available at \href{https://github.com/YubaoZhao/BranPO}{code}.

Seeing, Hearing, and Knowing Together: Multimodal Strategies in Deepfake Videos Detection

As deepfake videos become increasingly difficult for people to recognise, understanding the strategies humans use is key to designing effective media literacy interventions. We conducted a study with 195 participants between the ages of 21 and 40, who judged real and deepfake videos, rated their confidence, and reported the cues they relied on across visual, audio, and knowledge strategies. Participants were more accurate with real videos than with deepfakes and showed lower expected calibration error for real content. Through association rule mining, we identified cue combinations that shaped performance. Visual appearance, vocal, and intuition often co-occurred for successful identifications, which highlights the importance of multimodal approaches in human detection. Our findings show which cues help or hinder detection and suggest directions for designing media literacy tools that guide effective cue use. Building on these insights can help people improve their identification skills and become more resilient to deceptive digital media.

Uncovering Latent Communication Patterns in Brain Networks via Adaptive Flow Routing

Unraveling how macroscopic cognitive phenotypes emerge from microscopic neuronal connectivity remains one of the core pursuits of neuroscience. To this end, researchers typically leverage multi-modal information from structural connectivity (SC) and functional connectivity (FC) to complete downstream tasks. Recent methodologies explore the intricate coupling mechanisms between SC and FC, attempting to fuse their representations at the regional level. However, lacking fundamental neuroscientific insight, these approaches fail to uncover the latent interactions between neural regions underlying these connectomes, and thus cannot explain why SC and FC exhibit dynamic states of both coupling and heterogeneity. In this paper, we formulate multi-modal fusion through the lens of neural communication dynamics and propose the Adaptive Flow Routing Network (AFR-Net), a physics-informed framework that models how structural constraints (SC) give rise to functional communication patterns (FC), enabling interpretable discovery of critical neural pathways. Extensive experiments demonstrate that AFR-Net significantly outperforms state-of-the-art baselines. The code is available at https://anonymous.4open.science/r/DIAL-F0D1.

Post-LayerNorm Is Back: Stable, ExpressivE, and Deep

Large language model (LLM) scaling is hitting a wall. Widening models yields diminishing returns, and extending context length does not improve fundamental expressivity. In contrast, depth scaling offers theoretically superior expressivity, yet current Transformer architectures struggle to train reliably at extreme depths. We revisit the Post-LayerNorm (Post-LN) formulation, whose instability at scale caused its replacement by Pre-LN in modern LLMs. We show that the central failure mode of Post-LN arises from the ResNet-style residual pathway, which introduces gradient vanishing in deep networks. We present Keel, a Post-LN Transformer that replaces this residual path with a Highway-style connection. This modification preserves the gradient flow through the residual branch, preventing signal vanishing from the top layers to the bottom. Unlike prior methods, Keel enables stable training at extreme depths without requiring specialized initialization or complex optimization tricks. Keel trains robustly at depths exceeding 1000 layers and consistently improves perplexity and depth-scaling characteristics over Pre-LN. These findings indicate that Post-LN, when paired with a Highway-style connection, provides a simple and effective foundation for building deeply scalable LLMs, opening the possibility for future infinite-depth architectures.

The Shadow Self: Intrinsic Value Misalignment in Large Language Model Agents

Large language model (LLM) agents with extended autonomy unlock new capabilities, but also introduce heightened challenges for LLM safety. In particular, an LLM agent may pursue objectives that deviate from human values and ethical norms, a risk known as value misalignment. Existing evaluations primarily focus on responses to explicit harmful input or robustness against system failure, while value misalignment in realistic, fully benign, and agentic settings remains largely underexplored. To fill this gap, we first formalize the Loss-of-Control risk and identify the previously underexamined Intrinsic Value Misalignment (Intrinsic VM). We then introduce IMPRESS (Intrinsic Value Misalignment Probes in REalistic Scenario Set), a scenario-driven framework for systematically assessing this risk. Following our framework, we construct benchmarks composed of realistic, fully benign, and contextualized scenarios, using a multi-stage LLM generation pipeline with rigorous quality control. We evaluate Intrinsic VM on 21 state-of-the-art LLM agents and find that it is a common and broadly observed safety risk across models. Moreover, the misalignment rates vary by motives, risk types, model scales, and architectures. While decoding strategies and hyperparameters exhibit only marginal influence, contextualization and framing mechanisms significantly shape misalignment behaviors. Finally, we conduct human verification to validate our automated judgments and assess existing mitigation strategies, such as safety prompting and guardrails, which show instability or limited effectiveness. We further demonstrate key use cases of IMPRESS across the AI Ecosystem. Our code and benchmark will be publicly released upon acceptance.

GUI-Eyes: Tool-Augmented Perception for Visual Grounding in GUI Agents

Recent advances in vision-language models (VLMs) and reinforcement learning (RL) have driven progress in GUI automation. However, most existing methods rely on static, one-shot visual inputs and passive perception, lacking the ability to adaptively determine when, whether, and how to observe the interface. We present GUI-Eyes, a reinforcement learning framework for active visual perception in GUI tasks. To acquire more informative observations, the agent learns to make strategic decisions on both whether and how to invoke visual tools, such as cropping or zooming, within a two-stage reasoning process. To support this behavior, we introduce a progressive perception strategy that decomposes decision-making into coarse exploration and fine-grained grounding, coordinated by a two-level policy. In addition, we design a spatially continuous reward function tailored to tool usage, which integrates both location proximity and region overlap to provide dense supervision and alleviate the reward sparsity common in GUI environments. On the ScreenSpot-Pro benchmark, GUI-Eyes-3B achieves 44.8% grounding accuracy using only 3k labeled samples, significantly outperforming both supervised and RL-based baselines. These results highlight that tool-aware active perception, enabled by staged policy reasoning and fine-grained reward feedback, is critical for building robust and data-efficient GUI agents.

HGATSolver: A Heterogeneous Graph Attention Solver for Fluid-Structure Interaction

Fluid-structure interaction (FSI) systems involve distinct physical domains, fluid and solid, governed by different partial differential equations and coupled at a dynamic interface. While learning-based solvers offer a promising alternative to costly numerical simulations, existing methods struggle to capture the heterogeneous dynamics of FSI within a unified framework. This challenge is further exacerbated by inconsistencies in response across domains due to interface coupling and by disparities in learning difficulty across fluid and solid regions, leading to instability during prediction. To address these challenges, we propose the Heterogeneous Graph Attention Solver (HGATSolver). HGATSolver encodes the system as a heterogeneous graph, embedding physical structure directly into the model via distinct node and edge types for fluid, solid, and interface regions. This enables specialized message-passing mechanisms tailored to each physical domain. To stabilize explicit time stepping, we introduce a novel physics-conditioned gating mechanism that serves as a learnable, adaptive relaxation factor. Furthermore, an Inter-domain Gradient-Balancing Loss dynamically balances the optimization objectives across domains based on predictive uncertainty. Extensive experiments on two constructed FSI benchmarks and a public dataset demonstrate that HGATSolver achieves state-of-the-art performance, establishing an effective framework for surrogate modeling of coupled multi-physics systems.

Monkey Jump : MoE-Style PEFT for Efficient Multi-Task Learning

Mixture-of-experts variants of parameter-efficient fine-tuning enable per-token specialization, but they introduce additional trainable routers and expert parameters, increasing memory usage and training cost. This undermines the core goal of parameter-efficient fine-tuning. We propose Monkey Jump, a method that brings mixture-of-experts-style specialization to parameter-efficient fine-tuning without introducing extra trainable parameters for experts or routers. Instead of adding new adapters as experts, Monkey Jump treats the adapters already present in each Transformer block (such as query, key, value, up, and down projections) as implicit experts and routes tokens among them. Routing is performed using k-means clustering with exponentially moving averaged cluster centers, requiring no gradients and no learned parameters. We theoretically show that token-wise routing increases expressivity and can outperform shared adapters by avoiding cancellation effects. Across multi-task experiments covering 14 text, 14 image, and 19 video benchmarks, Monkey Jump achieves competitive performance with mixture-of-experts-based parameter-efficient fine-tuning methods while using 7 to 29 times fewer trainable parameters, up to 48 percent lower memory consumption, and 1.5 to 2 times faster training. Monkey Jump is architecture-agnostic and can be applied to any adapter-based parameter-efficient fine-tuning method.

Learning from Prompt itself: the Hierarchical Attribution Prompt Optimization

Optimization is fundamental across numerous disciplines, typically following an iterative process of refining an initial solution to enhance performance. This principle is equally critical in prompt engineering, where designing effective prompts for large language models constitutes a complex optimization challenge. A structured optimization approach requires automated or semi-automated procedures to develop improved prompts, thereby reducing manual effort, improving performance, and yielding an interpretable process. However, current prompt optimization methods often induce prompt drift, where new prompts fix prior failures but impair performance on previously successful tasks. Additionally, generating prompts from scratch can compromise interpretability. To address these limitations, this study proposes the Hierarchical Attribution Prompt Optimization (HAPO) framework, which introduces three innovations: (1) a dynamic attribution mechanism targeting error patterns in training data and prompting history, (2) semantic-unit optimization for editing functional prompt segments, and (3) multimodal-friendly progression supporting both end-to-end LLM and LLM-MLLM workflows. Applied in contexts like single/multi-image QA (e.g., OCRV2) and complex task analysis (e.g., BBH), HAPO demonstrates enhanced optimization efficiency, outperforming comparable automated prompt optimization methods and establishing an extensible paradigm for scalable prompt engineering.

Communication-Efficient Federated AUC Maximization with Cyclic Client Participation

Federated AUC maximization is a powerful approach for learning from imbalanced data in federated learning (FL). However, existing methods typically assume full client availability, which is rarely practical. In real-world FL systems, clients often participate in a cyclic manner: joining training according to a fixed, repeating schedule. This setting poses unique optimization challenges for the non-decomposable AUC objective. This paper addresses these challenges by developing and analyzing communication-efficient algorithms for federated AUC maximization under cyclic client participation. We investigate two key settings: First, we study AUC maximization with a squared surrogate loss, which reformulates the problem as a nonconvex-strongly-concave minimax optimization. By leveraging the Polyak-Łojasiewicz (PL) condition, we establish a state-of-the-art communication complexity of $\widetilde{O}(1/ε^{1/2})$ and iteration complexity of $\widetilde{O}(1/ε)$. Second, we consider general pairwise AUC losses. We establish a communication complexity of $O(1/ε^3)$ and an iteration complexity of $O(1/ε^4)$. Further, under the PL condition, these bounds improve to communication complexity of $\widetilde{O}(1/ε^{1/2})$ and iteration complexity of $\widetilde{O}(1/ε)$. Extensive experiments on benchmark tasks in image classification, medical imaging, and fraud detection demonstrate the superior efficiency and effectiveness of our proposed methods.