REVIEW 39 cited by
AgentTuning: Enabling Generalized Agent Abilities for LLMs
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Signed reviews
read the original abstract
Open large language models (LLMs) with great performance in various tasks have significantly advanced the development of LLMs. However, they are far inferior to commercial models such as ChatGPT and GPT-4 when acting as agents to tackle complex tasks in the real world. These agent tasks employ LLMs as the central controller responsible for planning, memorization, and tool utilization, necessitating both fine-grained prompting methods and robust LLMs to achieve satisfactory performance. Though many prompting methods have been proposed to complete particular agent tasks, there is lack of research focusing on improving the agent capabilities of LLMs themselves without compromising their general abilities. In this work, we present AgentTuning, a simple and general method to enhance the agent abilities of LLMs while maintaining their general LLM capabilities. We construct AgentInstruct, a lightweight instruction-tuning dataset containing high-quality interaction trajectories. We employ a hybrid instruction-tuning strategy by combining AgentInstruct with open-source instructions from general domains. AgentTuning is used to instruction-tune the Llama 2 series, resulting in AgentLM. Our evaluations show that AgentTuning enables LLMs' agent capabilities without compromising general abilities. The AgentLM-70B is comparable to GPT-3.5-turbo on unseen agent tasks, demonstrating generalized agent capabilities. We open source the AgentInstruct and AgentLM-7B, 13B, and 70B models at https://github.com/THUDM/AgentTuning, serving open and powerful alternatives to commercial LLMs for agent tasks.
Forward citations
Cited by 39 Pith papers
-
Reclaim Evaluation: A Lossy Memory Is Worse Than an Empty One
Reclaim evaluation shows lossy memory in language models is never better than empty memory across eight models, with a source-first policy restoring correctability at fixed budget.
-
ISE: An Execution-Grounded Recipe for Multi-Turn OS-Agent Trajectories
ISE creates 23,132 execution-grounded multi-turn OS agent trajectories via intent simulation and live execution, improving agent performance on ClawEval from 19.3 to 37.7 pass@1 with Qwen3-8B.
-
FinReportBench: Measuring and Improving Institution-Grade Financial Report Generation
FinReportBench is a fine-grained, expert-grounded benchmark for institution-grade LLM financial report generation, and its skill-evolution method improves G1 and G2 scores across model families.
-
Eluna: An Agentic LLM System for Automating Warehouse Operations with Reasoning and Task Execution
A graph-guided multi-agent framework plus asymmetric episodic distillation lets fine-tuned 32B models match larger teachers on warehouse SOP execution and hit 94% expert agreement on ticket processing.
-
BVS: Bayesian Visual Search with Multimodal Large Language Model for Fine-grained Perception
BVS combines early-stop attention rollout priors with a scale-aware non-stationary kernel and GP-UCB to locate tiny objects in UHR images more accurately and with fewer MLLM queries than prior visual-search methods.
-
MMAT-1M: A Large Reasoning Dataset for Multimodal Agent Tuning
A new one-million-sample multimodal agent tuning dataset with GPT-4o-generated rationales, reflection, and tool/RAG calls is shown to improve fine-tuned models, though training/eval benchmark overlap is not addressed.
-
Your Agent Can Defend Itself against Backdoor Attacks
A two-level consistency defense detects backdoored LLM agents by matching thoughts to actions and reconstructed instructions to the user's instruction, reducing attack success rates on tested tasks.
-
Synthesis by Design: Controlled Data Generation via Structural Guidance
A structural code-intervention method generates new math problems with labeled intermediate steps and a harder benchmark, and fine-tuning on the data mostly improves LLM math performance.
-
LLM-ODDR: A Large Language Model Framework for Joint Order Dispatching and Driver Repositioning
LLM-ODDR uses prompted and fine-tuned large language models for joint order dispatching and driver repositioning and reports higher GMV and order response rate than eight baselines in a simulated Manhattan taxi environment.
-
ManiTaskGen: A Comprehensive Task Generator for Benchmarking and Improving Vision-Language Agents on Embodied Decision-Making
ManiTaskGen automatically generates diverse, feasible mobile manipulation tasks from any input scene, and uses them to benchmark and improve vision-language robot agents.
-
Training LLM-Based Agents with Synthetic Self-Reflected Trajectories and Partial Masking
A new agent-training method combining teacher-generated self-reflection corrections with partial masking of error steps improves open-source LLM agents on ALFWorld, WebShop, and SciWorld.
-
Divide and Conquer: Grounding LLMs as Efficient Decision-Making Agents via Offline Hierarchical Reinforcement Learning
A prompt-shared hierarchical offline RL pipeline for LLM agents improves long-horizon task scores on ScienceWorld and ALFWorld over non-hierarchical baselines.
-
USTBench: Benchmarking and Dissecting Spatiotemporal Reasoning of LLMs as Urban Agents
USTBench is the first benchmark that decomposes urban spatiotemporal reasoning into understanding, forecasting, planning, and reflection, and shows LLMs struggle most with planning and reflection.
-
Effective Reinforcement Learning for Reasoning in Language Models
DASH uses large-batch sampling plus filtering of low-advantage examples to cut RL training time for small reasoning models by about 83% with no measured accuracy loss.
-
DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation
DynamicRAG trains a reranker as an RL agent, using the generator's answer quality as reward to dynamically choose how many and which retrieved documents to pass forward.
-
Towards Efficient Online Tuning of VLM Agents via Counterfactual Soft Reinforcement Learning
CoSo uses counterfactual token sensitivity from a learned surrogate to reweight entropy regularization in RL, improving online fine-tuning of VLM agents across three task families.
-
Exploring Expert Failures Improves LLM Agent Tuning
EEF improves LLM agent fine-tuning by extracting beneficial action segments from failed expert trajectories via rollouts from intermediate states, reporting a 62% WebShop win rate and a ScienceWorld score above 81.
-
Process Reward Models for LLM Agents: Practical Framework and Directions
A lightweight actor-critic loop trains 3B LLM agents to surpass GPT-4o on ALFWorld by learning step-level reward models from rollouts or demonstrations.
-
InSTA: Towards Internet-Scale Training For Agents
Automated LLM task generation, agent execution, and judge filtering at 150k-site scale lets a 1.7B model match much larger web agents.
-
Learn-by-interact: A Data-Centric Framework for Self-Adaptive Agents in Realistic Environments
A data-centric method that relabels agent trajectories with new instructions, called backward construction, improves LLM agent performance on web, code, desktop, and data-science tasks without human labeling.
-
Disentangling Exploration of Large Language Models by Optimal Exploitation
Exploration by LLM agents can be measured separately from exploitation using an optimal exploitation oracle, and most models explore poorly, with exploration performance correlated to reasoning ability.
-
AgentRefine: Enhancing Agent Generalization through Refinement Tuning
Training LLM agents on synthetic trajectories that include mistake-and-refinement steps improves generalization to unseen agent environments.
-
Aviary: training language agents on challenging scientific tasks
A small open-source LLM trained in the new Aviary environments with expert iteration and majority voting matches or exceeds a frontier LLM agent on SeqQA and LitQA2 at far lower inference cost.
-
Disentangling Reasoning Tokens and Boilerplate Tokens For Language Model Fine-tuning
A shuffle-based token classifier plus group-level loss reweighting improves supervised fine-tuning of LLM agents on tool-use benchmarks.
-
Towards Adaptive Mechanism Activation in Language Agent
ALAMA trains Llama-3-8B to choose among five agent mechanisms per task via self-exploration and KTO-based preference learning, improving accuracy over fixed-mechanism baselines on math and QA tasks.
-
TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents
Turn-level rewards from TD changes in a frozen reference model's gold-answer log-probability improve long-horizon search-agent RL on closed- and open-web benchmarks.
-
RoboGPT-R1: Enhancing Robot Task Planning with Reinforcement Learning
A 3B VLM trained with SFT plus GRPO and an LCS-based reward reaches 55.3% on EmbodiedBench's EB-ALFRED, beating GPT-4o-mini and the 7B REBP planner.
-
S3LoRA: Safe Spectral Sharpness-Guided Pruning in Adaptation of Agent Planner
S3LoRA prunes LoRA layers with the sharpest spectral update concentration to improve safety in fine-tuned LLM agents without needing base models or extra data.
-
DatasetAgent: A Novel Multi-Agent System for Auto-Constructing Datasets from Real-World Images
DatasetAgent is an LLM-powered multi-agent pipeline that automatically constructs image classification, detection, and segmentation datasets from web images, with modest downstream gains shown but weak experimental controls.
-
Agent-RewardBench: Towards a Unified Benchmark for Reward Modeling across Perception, Planning, and Safety in Real-World Multimodal Agents
A new step-level reward-modeling benchmark for multimodal agents shows current MLLMs reach at most 61.6 percent accuracy, and benchmark score correlates strongly (r=0.981 across five models) with downstream A* search ...
-
CRITICTOOL: Evaluating Self-Critique Capabilities of Large Language Models in Tool-Calling Error Scenarios
CRITICTOOL, a benchmark of 2,740 tool-calling error scenarios built from BFCL and T-Eval with GPT-4o-based error injection, finds most LLMs rarely recover from tool-use errors, with GPT-4o best at 69.01 overall and to...
-
Automated Skill Discovery for Language Agents through Exploration and Iterative Feedback
EXIF repeatedly has a teacher agent explore an environment, relabel the exploration as tasks, train a student agent on it, and use the student's failures to guide the next round, improving 7B-8B agents in Webshop and Crafter.
-
OWL: Optimized Workforce Learning for General Multi-Agent Assistance in Real-World Task Automation
A hierarchical multi-agent framework with a domain-agnostic planner trained by SFT and DPO achieves 69.70% on GAIA and lifts a Qwen2.5-32B planner by 16.37 points.
-
Digi-Q: Learning Q-Value Functions for Training Device-Control Agents
An offline RL method learns a Q-function from frozen VLM features and extracts a device-control policy by imitating the best of several actions ranked by that Q-function.
-
The Dawn of GUI Agent: A Preliminary Case Study with Claude 3.5 Computer Use
A qualitative case study of Claude 3.5 Computer Use across 20 desktop tasks shows broad but shallow capability, with notable failures in precise text selection and self-verification.
-
SPA-RL: Reinforcing LLM Agents via Stepwise Progress Attribution
SPA-RL attributes a single final task reward to individual agent steps and uses the attributions plus an executability signal as dense rewards for PPO, yielding modest success-rate gains on WebShop, ALFWorld, and VirtualHome.
-
Implicit Communication of Contextual Information in Human-Robot Collaboration
A thesis proposal claiming that robots that interpret linguistic implicatures improve perceived team performance and trust, and outlining future studies on robot-generated implicit cues and a multi-LLM system.
-
Training Agents with Weakly Supervised Feedback from Large Language Models
Iteratively fine-tuning LLM agents on trajectories selected by a critic LLM yields API-Bank accuracy close to GPT-4 using much smaller models.
-
Token-Operations-Oriented Inference Optimization Techniques for Large Models
The paper introduces a four-layer technical architecture for token-operations-oriented inference optimization in large models and reviews key technologies and industry status at each layer.
Discussion (0). Continue with ORCID to comment.