DuMate-DeepResearch: An Auditable Multi-Agent System with Recursive Search and Rubric-Grounded Reasoning

Deep Research (DR) has emerged as a new agentic paradigm to tackle complex, open-ended research tasks, demanding systems that can iteratively frame problems, acquire evidence, verify sources, and synthesize long-form reports. In practice, however, current DR systems are constrained by four interrelated limitations: long-horizon planning over an underspecified scope, the bottleneck of decomposing and scheduling such tasks within a single agent, hallucination risk in long-form synthesis, and limited process auditability. This technical report presents DuMate-DeepResearch, a multi-agent DR framework built on the Qianfan Agent Foundry. The framework decouples the Agent Core, which handles task understanding, planning, and scheduling, from an extensible Tool Ecosystem for retrieval, evidence acquisition, and report rendering, making every intermediate decision and tool invocation explicitly traceable. Building on this infrastructure, DuMate-DeepResearch further introduces three mechanisms: (i) a graph-based dynamic planning strategy expands the research roadmap coarse-to-fine and continuously revises it through reflection, re-planning, backtracking, and parallel branching; (ii) a recursive two-level execution design delegates each complex search sub-task to an inner Search Agent that runs its own planning loop, isolating noisy retrieval and stabilizing long-horizon execution; (iii) a rubric-based test-time optimization mechanism dynamically generates task-specific quality criteria and uses them as live reasoning scaffolds for evidence-grounded synthesis and adaptive stopping. Across two deep research benchmarks, DuMate-DeepResearch establishes new state-of-the-art results: the best overall score (58.03%) on DeepResearch Bench, and the best overall score (61.95%) on DeepResearch Bench II while ranking first in information recall and analysis.

When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents

Existing benchmarks evaluate Tool-Integrated Reasoning (TIR) in LLMs on idealized ''happy paths'', largely overlooking real-world tool failures. We introduce ToolMaze, a benchmark for dynamic path discovery and error recovery in TIR agents. To separate systematic replanning from blind trial-and-error, ToolMaze adopts a two-dimensional design: DAG-based topological complexity and a $2 \times 2$ taxonomy of tool perturbations (explicit/implicit, transient/permanent). Evaluations show that perturbations degrade performance across nearly all models, with the sharpest drops under implicit semantic failures. Driven by systemic over-trust in corrupted outputs, Perturbation Recovery Rate (PRR) plummets by around 37\% in these scenarios, while complex topologies trap agents in futile trial-and-error loops. Crucially, agentic fault-tolerance improves with model scale $3.66\times$ slower than basic task execution, highlighting dynamic replanning as a distinct bottleneck unaddressed by model scaling or prompting. Data and code are available at https://github.com/Zhudongsheng75/ToolMaze.

EndPrompt: Efficient Long-Context Extension via Terminal Anchoring

Extending the context window of large language models typically requires training on sequences at the target length, incurring quadratic memory and computational costs that make long-context adaptation expensive and difficult to reproduce. We propose EndPrompt, a method that achieves effective context extension using only short training sequences. The core insight is that exposing a model to long-range relative positional distances does not require constructing full-length inputs: we preserve the original short context as an intact first segment and append a brief terminal prompt as a second segment, assigning it positional indices near the target context length. This two-segment construction introduces both local and long-range relative distances within a short physical sequence while maintaining the semantic continuity of the training text--a property absent in chunk-based simulation approaches that split contiguous context. We provide a theoretical analysis grounded in Rotary Position Embedding and the Bernstein inequality, showing that position interpolation induces a rigorous smoothness constraint over the attention function, with shared Transformer parameters further suppressing unstable extrapolation to unobserved intermediate distances. Applied to LLaMA-family models extending the context window from 8K to 64K, EndPrompt achieves an average RULER score of 76.03 and the highest average on LongBench, surpassing LCEG (72.24), LongLoRA (72.95), and full-length fine-tuning (69.23) while requiring substantially less computation. These results demonstrate that long-context generalization can be induced from sparse positional supervision, challenging the prevailing assumption that dense long-sequence training is necessary for reliable context-window extension. The code is available at https://github.com/clx1415926/EndPrompt.

LLM-based Listwise Reranking under the Effect of Positional Bias

LLM-based listwise passage reranking has attracted attention for its effectiveness in ranking candidate passages. However, these models suffer from positional bias, where passages positioned towards the end of the input are less likely to be moved to top positions in the ranking. We hypothesize that there are two primary sources of positional bias: (1) architectural bias inherent in LLMs and (2) the imbalanced positioning of relevant documents. To address this, we propose DebiasFirst, a method that integrates positional calibration and position-aware data augmentation during fine-tuning. Positional calibration uses inverse propensity scoring to adjust for positional bias by re-weighting the contributions of different positions in the loss function when training. Position-aware augmentation augments training data to ensure that each passage appears equally across varied positions in the input list. This approach markedly enhances both effectiveness and robustness to the original ranking across diverse first-stage retrievers, reducing the dependence of NDCG@10 performance on the position of relevant documents. DebiasFirst also complements the inference-stage debiasing methods, offering a practical solution for mitigating positional bias in reranking.

How to Utilize Complementary Vision-Text Information for 2D Structure Understanding

LLMs typically linearize 2D tables into 1D sequences to fit their autoregressive architecture, which weakens row-column adjacency and other layout cues. In contrast, purely visual encoders can capture spatial cues, yet often struggle to preserve exact cell text. Our analysis reveals that these two modalities provide highly distinct information to LLMs and exhibit strong complementarity. However, direct concatenation and other fusion methods yield limited gains and frequently introduce cross-modal interference. To address this issue, we propose DiVA-Former, a lightweight architecture designed to effectively integrate vision and text information. DiVA-Former leverages visual tokens as dynamic queries to distill long textual sequences into digest vectors, thereby effectively exploiting complementary vision--text information. Evaluated across 13 table benchmarks, DiVA-Former improves upon the pure-text baseline by 23.9\% and achieves consistent gains over existing baselines using visual inputs, textual inputs, or a combination of both.

AdaFuse: Accelerating Dynamic Adapter Inference via Token-Level Pre-Gating and Fused Kernel Optimization

The integration of dynamic, sparse structures like Mixture-of-Experts (MoE) with parameter-efficient adapters (e.g., LoRA) is a powerful technique for enhancing Large Language Models (LLMs). However, this architectural enhancement comes at a steep cost: despite minimal increases in computational load, the inference latency often skyrockets, leading to decoding speeds slowing by over 2.5 times. Through a fine-grained performance analysis, we pinpoint the primary bottleneck not in the computation itself, but in the severe overhead from fragmented, sequential CUDA kernel launches required for conventional dynamic routing. To address this challenge, we introduce AdaFuse, a framework built on a tight co-design between the algorithm and the underlying hardware system to enable efficient dynamic adapter execution. Departing from conventional layer-wise or block-wise routing, AdaFuse employs a token-level pre-gating strategy, which makes a single, global routing decision for all adapter layers before a token is processed. This "decide-once, apply-everywhere" approach effectively staticizes the execution path for each token, creating an opportunity for holistic optimization. We capitalize on this by developing a custom CUDA kernel that performs a fused switching operation, merging the parameters of all selected LoRA adapters into the backbone model in a single, efficient pass. Experimental results on popular open-source LLMs show that AdaFuse achieves accuracy on par with state-of-the-art dynamic adapters while drastically cutting decoding latency by a factor of over 2.4x, thereby bridging the gap between model capability and inference efficiency.

Model Editing for New Document Integration in Generative Information Retrieval

Generative retrieval (GR) reformulates the Information Retrieval (IR) task as the generation of document identifiers (docIDs). Despite its promise, existing GR models exhibit poor generalization to newly added documents, often failing to generate the correct docIDs. While incremental training offers a straightforward remedy, it is computationally expensive, resource-intensive, and prone to catastrophic forgetting, thereby limiting the scalability and practicality of GR. In this paper, we identify the core bottleneck as the decoder's ability to map hidden states to the correct docIDs of newly added documents. Model editing, which enables targeted parameter modifications for docID mapping, represents a promising solution. However, applying model editing to current GR models is not trivial, which is severely hindered by indistinguishable edit vectors across queries, due to the high overlap of shared docIDs in retrieval results. To address this, we propose DOME (docID-oriented model editing), a novel method that effectively and efficiently adapts GR models to unseen documents. DOME comprises three stages: (1) identification of critical layers, (2) optimization of edit vectors, and (3) construction and application of updates. At its core, DOME employs a hybrid-label adaptive training strategy that learns discriminative edit vectors by combining soft labels, which preserve query-specific semantics for distinguishable updates, with hard labels that enforce precise mapping modifications. Experiments on widely used benchmarks, including NQ and MS MARCO, show that our method significantly improves retrieval performance on new documents while maintaining effectiveness on the original collection. Moreover, DOME achieves this with only about 60% of the training time required by incremental training, considerably reducing computational cost and enabling efficient, frequent model updates.

DeepResearch-9K: A Challenging Benchmark Dataset of Deep-Research Agent

Deep-research agents are capable of executing multi-step web exploration, targeted retrieval, and sophisticated question answering. Despite their powerful capabilities, deep-research agents face two critical bottlenecks: (1) the lack of large-scale, challenging datasets with real-world difficulty, and (2) the absence of accessible, open-source frameworks for data synthesis and agent training. To bridge these gaps, we first construct DeepResearch-9K, a large-scale challenging dataset specifically designed for deep-research scenarios built from open-source multi-hop question-answering (QA) datasets via a low-cost autonomous pipeline. Notably, it consists of (1) 9000 questions spanning three difficulty levels from L1 to L3 (2) high-quality search trajectories with reasoning chains from Tongyi-DeepResearch-30B-A3B, a state-of-the-art deep-research agent, and (3) verifiable answers. Furthermore, we develop an open-source training framework DeepResearch-R1 that supports (1) multi-turn web interactions, (2) different reinforcement learning (RL) approaches, and (3) different reward models such as rule-based outcome reward and LLM-as-judge feedback. Finally, empirical results demonstrate that agents trained on DeepResearch-9K under our DeepResearch-R1 achieve state-of-the-art results on challenging deep-research benchmarks. We release the DeepResearch-9K dataset on https://huggingface.co/datasets/artillerywu/DeepResearch-9K and the code of DeepResearch-R1 on https://github.com/Applied-Machine-Learning-Lab/DeepResearch-R1.

AgentSkiller: Scaling Generalist Agent Intelligence through Semantically Integrated Cross-Domain Data Synthesis

Large Language Model agents demonstrate potential in solving real-world problems via tools, yet generalist intelligence is bottlenecked by scarce high-quality, long-horizon data. Existing methods collect privacy-constrained API logs or generate scripted interactions lacking diversity, which struggle to produce data requisite for scaling capabilities. We propose AgentSkiller, a fully automated framework synthesizing multi-turn interaction data across realistic, semantically linked domains. It employs a DAG-based architecture with explicit state transitions to ensure determinism and recoverability. The pipeline builds a domain ontology and Person-Centric Entity Graph, defines tool interfaces via Service Blueprints for Model Context Protocol servers, and populates environments with consistent databases and strict Domain Policies. A cross-domain fusion mechanism links services to simulate complex tasks. Finally, the pipeline creates user tasks by verifying solution paths, filtering via execution-based validation, and generating queries using a Persona-based Simulator for automated rollout. This produces reliable environments with clear state changes. To demonstrate effectiveness, we synthesized $\approx$ 11K interaction samples; experimental results indicate that models trained on this dataset achieve significant improvements on function calling over baselines, particularly in larger parameter regimes.

Not All Preferences Are Created Equal: Stability-Aware and Gradient-Efficient Alignment for Reasoning Models

Preference-based alignment is pivotal for training large reasoning models; however, standard methods like Direct Preference Optimization (DPO) typically treat all preference pairs uniformly, overlooking the evolving utility of training instances. This static approach often leads to inefficient or unstable optimization, as it wastes computation on trivial pairs with negligible gradients and suffers from noise induced by samples near uncertain decision boundaries. Facing these challenges, we propose SAGE (Stability-Aware Gradient Efficiency), a dynamic framework designed to enhance alignment reliability by maximizing the Signal-to-Noise Ratio of policy updates. Concretely, SAGE integrates a coarse-grained curriculum mechanism that refreshes candidate pools based on model competence with a fine-grained, stability-aware scoring function that prioritizes informative, confident errors while filtering out unstable samples. Experiments on multiple mathematical reasoning benchmarks demonstrate that SAGE significantly accelerates convergence and outperforms static baselines, highlighting the critical role of policy-aware, stability-conscious data selection in reasoning alignment.