Shengji Tang

Peng Ye, Shengji Tang, Baopu Li et al. · deepmind

Residual networks have shown great success and become indispensable in today's deep models. In this work, we aim to re-investigate the training process of residual networks from a novel social psychology perspective of loafing, and further propose a new training strategy to strengthen the performance of residual networks. As residual networks can be viewed as ensembles of relatively shallow networks (i.e., \textit{unraveled view}) in prior works, we also start from such view and consider that the final performance of a residual network is co-determined by a group of sub-networks. Inspired by the social loafing problem of social psychology, we find that residual networks invariably suffer from similar problem, where sub-networks in a residual network are prone to exert less effort when working as part of the group compared to working alone. We define this previously overlooked problem as \textit{network loafing}. As social loafing will ultimately cause the low individual productivity and the reduced overall performance, network loafing will also hinder the performance of a given residual network and its sub-networks. Referring to the solutions of social psychology, we propose \textit{stimulative training}, which randomly samples a residual sub-network and calculates the KL-divergence loss between the sampled sub-network and the given residual network, to act as extra supervision for sub-networks and make the overall goal consistent. Comprehensive empirical results and theoretical analyses verify that stimulative training can well handle the loafing problem, and improve the performance of a residual network by improving the performance of its sub-networks. The code is available at https://github.com/Sunshine-Ye/NIPS22-ST .

Shengji Tang, Peng Ye, Baopu Li et al.

Recent research understands the residual networks from a new perspective of the implicit ensemble model. From this view, previous methods such as stochastic depth and stimulative training have further improved the performance of the residual network by sampling and training of its subnets. However, they both use the same supervision for all subnets of different capacities and neglect the valuable knowledge generated by subnets during training. In this manuscript, we mitigate the significant knowledge distillation gap caused by using the same kind of supervision and advocate leveraging the subnets to provide diverse knowledge. Based on this motivation, we propose a group knowledge based training framework for boosting the performance of residual networks. Specifically, we implicitly divide all subnets into hierarchical groups by subnet-in-subnet sampling, aggregate the knowledge of different subnets in each group during training, and exploit upper-level group knowledge to supervise lower-level subnet groups. Meanwhile, We also develop a subnet sampling strategy that naturally samples larger subnets, which are found to be more helpful than smaller subnets in boosting performance for hierarchical groups. Compared with typical subnet training and other methods, our method achieves the best efficiency and performance trade-offs on multiple datasets and network structures. The code is at https://github.com/tsj-001/AAAI24-GKT.

Hao Li, Yiqun Zhang, Zhaoyan Guo et al.

Large language model (LLM) routing assigns each query to the most suitable model from an ensemble. We introduce LLMRouterBench, a large-scale benchmark and unified framework for LLM routing. It comprises over 400K instances from 21 datasets and 33 models. Moreover, it provides comprehensive metrics for both performance-oriented routing and performance-cost trade-off routing, and integrates 10 representative routing baselines. Using LLMRouterBench, we systematically re-evaluate the field. While confirming strong model complementarity-the central premise of LLM routing-we find that many routing methods exhibit similar performance under unified evaluation, and several recent approaches, including commercial routers, fail to reliably outperform a simple baseline. Meanwhile, a substantial gap remains to the Oracle, driven primarily by persistent model-recall failures. We further show that backbone embedding models have limited impact, that larger ensembles exhibit diminishing returns compared to careful model curation, and that the benchmark also enables latency-aware analysis. All code and data are available at https://github.com/ynulihao/LLMRouterBench.

Wenzhen Yuan, Wutao Xiong, Fanchen Yu et al.

Multi-agent systems (MAS) demonstrate clear advantages in tackling complex problems by coordinating diverse agents and external tools. However, most existing orchestration methods rely on static workflows or serial agent scheduling, and are further constrained by heterogeneous interface protocols between tools and agents. This leads to high system complexity and poor extensibility. To mitigate these issues, we propose Agent-as-Tool, a unified parallel orchestration paradigm that abstracts both agents and tools into a standardized, learnable action space with protocol normalization and explicit state feedback. Building on this paradigm, we train a lightweight orchestrator, ParaManager, which decouples planning decisions from subtask solving, enabling state-aware parallel subtask decomposition, delegation, and asynchronous execution. For training, we adopt a two-stage ParaManager training pipeline. It improves robustness by incorporating supervised fine-tuning (SFT) trajectories equipped with recovery mechanisms, and further applies reinforcement learning (RL) to achieve an optimal balance among task success, protocol compliance, diversity, and reasoning efficiency. Experiments show that ParaManager achieves strong performance across multiple benchmarks and exhibits robust generalization under unseen model pools.

Zhende Song, Shengji Tang, Peng Ye et al.

Test-time scaling (TTS) has emerged as a promising, training-free approach for enhancing large language model (LLM) performance. However, the efficacy of existing methods, such as Best-of-N and Self-Consistency, is fundamentally constrained by the dominant single test-time scaling (STTS) paradigm, which relies on a single LLM agent interacting with a single reward model (SA-SR). Inspired by recent work showing that collective methods can surpass the performance ceiling of individual models, we introduce Collective Test-Time Scaling (CTTS). First, we systematically investigate three primary interaction paradigms of existing multiple models: single-agent-multi-reward (SA-MR), multi-agent-single-reward (MA-SR), and multi-agent-multi-reward (MA-MR). Extensive experiments reveal that the MA-MR paradigm is consistently superior. Based on this finding, we further propose CTTS-MM, a novel framework that operationalizes multi-agent and multi-reward collaboration. CTTS-MM integrates two key technical contributions: (1) for agent collaboration, an Agent Collaboration Search (ACS) that identifies the most effective combination of LLMs from a candidate pool; and (2) for reward model collaboration, a Mixture of Reward Models (MoR) strategy that leverages a Prior Reward model Ensemble Selection (PRES) algorithm to select the optimal ensemble. Evaluations across seven mainstream benchmarks demonstrate that CTTS-MM significantly outperforms leading STTS methods (+4.82% over Best-of-N) and surpasses even flagship proprietary LLMs (+7.06% over GPT-4.1) and open-source LLMs. These results highlight the substantial potential of collective scaling to push the frontier of LLM inference. Code will be released at https://github.com/magent4aci/CTTS-MM.

Shengji Tang, Weihao Lin, Jingqi Ye et al.

Large Language Models (LLMs) have rapidly advanced, with Gemini-3-Pro setting a new performance milestone. In this work, we explore collective intelligence as an alternative to monolithic scaling, and demonstrate that open-source LLMs' collaboration can surpass Gemini-3-Pro. We first revisit LLM routing and aggregation at scale and identify three key bottlenecks: (1) current train-free routers are limited by a query-based paradigm focusing solely on textual similarity; (2) recent aggregation methods remain largely static, failing to select appropriate aggregators for different tasks;(3) the complementarity of routing and aggregation remains underutilized. To address these problems, we introduce JiSi, a novel framework designed to release the full potential of LLMs' collaboration through three innovations: (1) Query-Response Mixed Routing capturing both semantic information and problem difficulty; (2) Support-Set-based Aggregator Selection jointly evaluating the aggregation and domain capacity of aggregators; (3) Adaptive Routing-Aggregation Switch dynamically leveraging the advantages of routing and aggregation. Comprehensive experiments on nine benchmarks demonstrate that JiSi can surpass Gemini-3-Pro with only 47% costs by orchestrating ten open-source LLMs, while outperforming mainstream baselines. It suggests that collective intelligence represents a novel path towards Artificial General Intelligence (AGI).

Wenzhen Yuan, Shengji Tang, Weihao Lin et al.

Reinforcement learning (RL) has significantly enhanced the reasoning capabilities of large language models (LLMs), but its reliance on expensive human-labeled data or complex reward models severely limits scalability. While existing self-feedback methods aim to address this problem, they are constrained by the capabilities of a single model, which can lead to overconfidence in incorrect answers, reward hacking, and even training collapse. To this end, we propose Reinforcement Learning from Coevolutionary Collective Feedback (RLCCF), a novel RL framework that enables multi-model collaborative evolution without external supervision. Specifically, RLCCF optimizes the ability of a model collective by maximizing its Collective Consistency (CC), which jointly trains a diverse ensemble of LLMs and provides reward signals by voting on collective outputs. Moreover, each model's vote is weighted by its Self-Consistency (SC) score, ensuring that more confident models contribute more to the collective decision. Benefiting from the diverse output distributions and complementary abilities of multiple LLMs, RLCCF enables the model collective to continuously enhance its reasoning ability through coevolution. Experiments on four mainstream open-source LLMs across four mathematical reasoning benchmarks demonstrate that our framework yields significant performance gains, achieving an average relative improvement of 16.72\% in accuracy. Notably, RLCCF not only improves the performance of individual models but also enhances the group's majority-voting accuracy by 4.51\%, demonstrating its ability to extend the collective capability boundary of the model collective.

Shengji Tang, Jianjian Cao, Weihao Lin et al.

This paper aims to demonstrate the potential and strengths of open-source collectives. It leads to a promising question: Can we harness multiple open-source LLMs to match or even beat the closed-source LLMs? To answer this, we propose SMACS, a scalable multi-agent collaboration system (MACS) framework with high performance. Specifically, for continuous integration of new LLMs and generalization to diverse questions, we first propose a Retrieval-based Prior Selection (RPS), which assigns a proxy performance score to each LLM to select the Top-k LLMs at the instance level for any given question. Then, we propose an Exploration-Exploitation-Driven Posterior Enhancement (EPE), encouraging the generation of diverse responses through prior dropping and selecting the high-quality response via a hybrid posterior score. Experiments on eight mainstream benchmarks validate the effectiveness of our SMACS: by integrating fifteen open-source LLMs, SMACS outperforms leading closed-source LLMs in 2025, e.g., Claude-3.7-Sonnet (+12.73%), GPT-4.1(+5.36%) and GPT-o3-mini(+5.28%) across multiple tasks. Remarkably, it even exceeds the average of best results of different datasets from both open-source LLMs (+2.86%) and closed-source LLMs (+2.04%), pushing the upper bound of intelligence. Code will be released at https://github.com/magent4aci/SMACS.

Xiangyuan Xue, Yifan Zhou, Zidong Wang et al.

Large language models (LLMs) are increasingly used as interactive agents, but optimizing them for long-horizon decision making remains difficult because current methods are largely purely reactive, which weakens both exploration and credit assignment over extended trajectories. In this work, we present Strategic Trajectory Abstraction (StraTA), a simple framework that introduces an explicit trajectory-level strategy into agentic reinforcement learning (RL). StraTA samples a compact strategy from the initial task state, conditions subsequent actions on that strategy, and trains strategy generation and action execution jointly with a hierarchical GRPO-style rollout design, further enhanced by diverse strategy rollout and critical self-judgment. Experiments on ALFWorld, WebShop, and SciWorld show that StraTA consistently improves both sample efficiency and final performance over strong baselines. StraTA reaches success rates of 93.1% on ALFWorld and 84.2% on WebShop. On SciWorld, StraTA attains a 63.5% overall score, outperforming frontier closed-source models.

Zhende Song, Chenchen Wang, Jiamu Sheng et al.

Recent large vision-language models (LVLMs) for video understanding are primarily fine-tuned with various videos scraped from online platforms. Existing datasets, such as ActivityNet, require considerable human labor for structuring and annotation before effectively utilized for tuning LVLMs. While current LVLMs are primarily trained on existing datasets in broad, general-purpose settings, adapting them to specific downstream scenarios remains challenging, as collecting and annotating task-specific videos is highly labor-intensive and time-consuming. To address this issue, we propose a three-stage framework named DreamFrame for automatically generating style-consistent keyframes and corresponding question-answer (QA) pairs to support LVLM instruction tuning. DreamFrame generates datasets in a movie-like manner. First, we utilize an LLM to generate structured movie plots including movie prior information (like overview and style), frame descriptions and plot-related QA pairs, with a story expansion strategy to mitigate context length limitations.Then, to ensure visual consistency across generated frames, we design a Style Immobilization Process which maintains consistent style through an embedding learning strategy. Finally, frame descriptions and style embeddings are integrated to produce coherent keyframes. Using DreamFrame, we construct a dataset comprising approximately 1k stylized keyframe-like videos and 100k diverse QA pairs. Extensive fine-tuned experiments on various LVLM architectures demonstrate the effectiveness of the proposed dataset. Furthermore, based on the proposed dataset, we fine-tune a new LVLM named DreamFrame-7B, which significantly surpasses the previous similar-sized LVLMs across different benchmarks.

Shengji Tang, Weihao Lin, Hancheng Ye et al.

Sparsification-based pruning has been an important category in model compression. Existing methods commonly set sparsity-inducing penalty terms to suppress the importance of dropped weights, which is regarded as the suppressed sparsification paradigm. However, this paradigm inactivates the dropped parts of networks causing capacity damage before pruning, thereby leading to performance degradation. To alleviate this issue, we first study and reveal the relative sparsity effect in emerging stimulative training and then propose a structured pruning framework, named STP, based on an enhanced sparsification paradigm which maintains the magnitude of dropped weights and enhances the expressivity of kept weights by self-distillation. Besides, to find an optimal architecture for the pruned network, we propose a multi-dimension architecture space and a knowledge distillation-guided exploration strategy. To reduce the huge capacity gap of distillation, we propose a subnet mutating expansion technique. Extensive experiments on various benchmarks indicate the effectiveness of STP. Specifically, without fine-tuning, our method consistently achieves superior performance at different budgets, especially under extremely aggressive pruning scenarios, e.g., remaining 95.11% Top-1 accuracy (72.43% in 76.15%) while reducing 85% FLOPs for ResNet-50 on ImageNet. Codes will be released soon.

Peng Ye, Tong He, Shengji Tang et al.

Residual networks have shown great success and become indispensable in recent deep neural network models. In this work, we aim to re-investigate the training process of residual networks from a novel social psychology perspective of loafing, and further propose a new training scheme as well as three improved strategies for boosting residual networks beyond their performance limits. Previous research has suggested that residual networks can be considered as ensembles of shallow networks, which implies that the final performance of a residual network is influenced by a group of subnetworks. We identify a previously overlooked problem that is analogous to social loafing, where subnetworks within a residual network are prone to exert less effort when working as part of a group compared to working alone. We define this problem as \textit{network loafing}. Similar to the decreased individual productivity and overall performance as demonstrated in society, network loafing inevitably causes sub-par performance. Inspired by solutions from social psychology, we first propose a novel training scheme called stimulative training, which randomly samples a residual subnetwork and calculates the KL divergence loss between the sampled subnetwork and the given residual network for extra supervision. In order to unleash the potential of stimulative training, we further propose three simple-yet-effective strategies, including a novel KL- loss that only aligns the network logits direction, random smaller inputs for subnetworks, and inter-stage sampling rules. Comprehensive experiments and analysis verify the effectiveness of stimulative training as well as its three improved strategies.