Source-linked AI summary
Scaling Agents via Continual Pre-training
Liangcai Su, Zhen Zhang, Guangyu Li, Zhuo Chen, Chenxi Wang, Maojia Song, Xinyu Wang, Kuan Li, Jialong Wu, Xuanzhong Chen, Zile Qiao, Zhongwang Zhang, Huifeng Yin, Shihao Cai, Runnan Fang, Zhengwei Tao, Wenbiao Yin, Chenxiong Qian, Yong Jiang, Pengjun Xie, Fei Huang, Jingren Zhou
TL;DR
Agentic post-training built on general-purpose foundations remains limited because models must acquire diverse agentic behaviors while aligning with demonstrations. The paper introduces Agentic CPT, develops AgentFounder through scalable action synthesis and staged training, and reports new SOTA results across 10 benchmarks with strong tool use. The authors note that Agentic CPT helps information retrieval more than knowledge-intensive tasks such as HLE.
Problem
Post-training methods built on general-purpose foundations underperform in agentic tasks because they must learn diverse agentic behaviors while aligning to expert demonstrations.
Method
Agentic CPT adds an intermediate training stage using scalable first- and higher-order action synthesis with a two-stage strategy to build agentic foundations before post-training.
Results
AgentFounder-30B achieves new SOTA results across 10 benchmarks, including 39.9% on BrowseComp-en, 43.3% on BrowseComp-zh, and 31.5% on HLE, while retaining strong tool-use ability.
Takeaways & Limitations
Agentic CPT supports strong deep-research performance and preserves general tool-use capabilities, indicating potential beyond retrieval reasoning toward broader agent tasks.
Takeaways & Limitations
Agentic CPT benefits information retrieval more than knowledge-intensive tasks such as HLE, which also require strong reasoning to use retrieved knowledge correctly.
Abstract
from arXiv · showhide
Large language models (LLMs) have evolved into agentic systems capable of autonomous tool use and multi-step reasoning for complex problem-solving. However, post-training approaches building upon general-purpose foundation models consistently underperform in agentic tasks, particularly in open-source implementations. We identify the root cause: the absence of robust agentic foundation models forces models during post-training to simultaneously learn diverse agentic behaviors while aligning them to expert demonstrations, thereby creating fundamental optimization tensions. To this end, we are the first to propose incorporating Agentic Continual Pre-training (Agentic CPT) into the deep research agents training pipeline to build powerful agentic foundational models. Based on this approach, we develop a deep research agent model named AgentFounder. We evaluate our AgentFounder-30B on 10 benchmarks and achieve state-of-the-art performance while retains strong tool-use ability, notably 39.9% on BrowseComp-en, 43.3% on BrowseComp-zh, and 31.5% Pass@1 on HLE.
1 Introdcution
The paper argues that general-purpose foundations create a bottleneck for agentic post-training and introduces Agentic CPT to build pre-aligned agentic foundations. AgentFounder combines scalable action synthesis with two-stage training and achieves strong results across 10 benchmarks while retaining tool-use ability.
- Motivation: Agentic alignment extends conventional alignment toward consistent behavior during multi-step reasoning, dynamic tool invocation, and complex environmental interaction.This reframes alignment for autonomous agents rather than static output preferences alone.
- Motivation: Existing post-training methods underperform because general-purpose foundations must learn diverse agentic behaviors while aligning to demonstrations, creating a critical bottleneck.The introduction contrasts open-source agentic models such as DeepSeek-V3.1 at 30.0 on BrowseComp with OpenAI Deep Research at 51.5.
- Contribution: Agentic CPT inserts an intermediate scaling layer that builds a pre-aligned agentic foundation model before downstream fine-tuning.Its data should span broad sources and diverse agentic behaviors rather than a single domain or memorized behavioral pattern.
- Method: AgentFounder is trained from Qwen3 models using First-order Action Synthesis, Higher-order Action Synthesis, and a two-stage training strategy.The synthesis pipeline is designed to support planning, reasoning, exploration, and multi-step decision-making.
- Results: 39.9% on BrowseComp-en, 43.3% on BrowseComp-zh, 72.8% on GAIA, 31.5% on HLE, and 73.0% on xbench-DeepSearch establish new SOTA results across 10 benchmarks.The model also retains strong general tool-use capabilities, and scaling training data produces steady average gains across benchmarks.
2 AgentFounder: Agentic Foundation Model
AgentFounder inserts Agentic Continual Pre-Training between pre-training and post-training to build agentic foundation models, using scalable synthetic data for tool use, reasoning, and decision-making. The pipeline combines broad data collection, staged training, first-order synthesis, and higher-order action synthesis to prepare models for downstream fine-tuning.
- Agentic training pipeline: Agentic CPT is added between pre-training and post-training to provide a pre-aligned agentic foundation model for downstream fine-tuning.The pipeline starts from Qwen pre-trained models and uses Agentic CPT as an additional training stage.
- Agentic training pipeline: 200B tokens at 32K context initialize tool invocation patterns and multi-step reasoning, followed by 100B curated tokens at 128K context for long-horizon planning.The two stages progressively refine agentic behaviors and understanding of complex action spaces.
- First-order Action Synthesis: Multi-style question synthesis converts entity-linked knowledge into factual, numerical, multi-hop, and synthesis tasks requiring retrieval, integration, and tool use.This transformation creates dynamic problem-solving contexts from static knowledge sources.
- Training data: Agentic CPT adapts domain-agnostic agentic capabilities by combining discarded trajectories, historical tool results, and public corpora spanning diverse scenarios.The data sources are intended to provide scalability and broad domain coverage.
- First-order Action Synthesis: Entity-anchored open-world memory reformulates continuously updated text into dense entity-indexed knowledge statements that preserve temporal, source, and stylistic information.The resulting memory is continuously expanded using search results and web-access outcomes.
- Higher-order Action Synthesis: Higher-order Action Synthesis expands alternative reasoning and tool-invocation paths at each trajectory step, transforming underused trajectories into decision-making training signals.The method explicitly shifts training from trajectory imitation toward step-wise decision-making without directly executing tools during synthesis.
3 Experiments
AgentFounder is evaluated across benchmark comparisons, post-training adaptability, training strategies, scaling behavior, efficiency, and tool-use patterns. The experiments report strong performance, consistent benefits from Agentic CPT, and adaptive tool use across task types.
- Performance Comparison: 31.5% on HLE makes AgentFounder-30B the first open-source model above 30 points, while it also leads reported models on Academic Browse and RACE Overall.
- Adaptability of the Agentic Base Model to Post-Training: Agentic CPT improves downstream post-training consistently, with average gains of 5.75%, 6.13%, and 6.45% across SFT-A, SFT-B, and SFT-C.
- Adaptability of the Agentic Base Model to Post-Training: Post-training data remains consequential: on BrowseComp-zh, SFT-B exceeds SFT-A and SFT-C by 8.5% and 3.0%, respectively.
- Impact of Training Strategies: The complete two-stage strategy improves Pass@1 by 3.3% and Pass@3 by 3.7% over single-stage training across three tasks.
- Scaling with Model Size: AgentFounder scales from 20.4% accuracy at 1B parameters to 48.9% at 30B, exceeding DeepSeek-V3.1 at 43.0% and Kimi-K2 at 29.6%.
- Training Process: CPT data scaling produces a total 8.0% gain from 54.2% to 62.2% across 315B tokens, while the best model lowers final SFT loss from 0.8656 to 0.7953.
- Tool Call Analysis: AgentFounder adapts tool-use intensity to task type, using dense long-tailed invocation patterns for complex research and conservative usage for structured navigation.
4 Related Work
Related work spans deep research agents, open-source agent development, continual pre-training, and emerging multi-agent inference paradigms. Existing continual pre-training research has mainly addressed non-tool tasks, leaving agentic capability integration largely unexplored.
- Deep research agents autonomously invoke tools and perform multi-step reasoning for research-level tasks and complex problem solving.
- Open-source agents have advanced across BrowseComp-en, BrowseComp-zh, Xbench-DeepSearch, and GAIA, while recent general models emphasize stronger agentic capabilities.
- Existing work focuses mainly on continual pre-training without tool calling or post-training, leaving direct integration of agentic reasoning and tool use into continual pre-training largely unexplored.
5 Conclusion
The paper introduces Agentic CPT as a training stage beyond traditional post-training and develops AgentFounder-30B through scalable synthesis and progressive training. AgentFounder-30B surpasses current closed-source models across 10 benchmarks.
- AgentFounder-30B uses Agentic CPT with scalable FAS and HAS synthesis plus an adapted two-stage training strategy, surpassing current closed-source models across 10 benchmarks.
- Agentic CPT adds planning and reasoning action synthesis without commercial API calls, while higher-order synthesis expands trajectories into multi-step decision-making paths.
- The proposed pipeline redefines agentic alignment training by embedding agentic capabilities beyond traditional post-training.
A Experimental Setup Details
AgentFounder uses five tools spanning web search, page extraction, code execution, scholarly retrieval, and file analysis. These tools support information gathering, computation, and multimodal document processing.
- AgentFounder uses Search, Visit, Python Interpreter, Google Scholar, and File Parser as its five tools.
- Search retrieves top-10 results for concurrent queries, while Visit extracts goal-relevant information from retrieved web pages.
- Python Interpreter executes code in a sandbox, Google Scholar retrieves academic literature, and File Parser converts diverse inputs into text for answering queries.
B.1 Quality Analysis of FAS data on Planning Action
The FAS quality analysis evaluates weakly supervised filtering of planning-action trajectories. Filtering removes problematic samples and substantially raises retained accuracy, while the accompanying figures document filtering performance and BrowseComp-en scaling.
- The experiment compares FAS planning-action quality before and after weak supervision using accuracy rates and error-type distributions.
- 82% retained trajectory accuracy follows filtering, up from 50%, after the filter removes 43.5% of problematic samples.
- Figure 9 presents filtering performance and representative low-quality outputs, while Figure 10 presents Pass@N scaling on BrowseComp-en.
B.2 Sampling Strategies and Performance Scaling
The study tests whether HAS preserves solution diversity under sampling and finds strong performance scaling as the number of sampled trajectories increases.
- The experiment examines whether HAS enhances behavioral diversity and enables effective scaling through sampling.
- The evaluation uses Pass@n on BrowseComp-en with sampling sizes from n=1 to 18 and increased temperature for multiple trajectories.This setup tests whether HAS preserves solution diversity after exposing the model to multiple potential actions during training.
- 75.8% Pass@16, up from 31.5% Pass@1, shows a +44.3 percentage point gain as sampling increases.Performance improves progressively across sampling sizes, with only a 1.16% improvement from Pass@16 to Pass@18.
B.3 Performance of GAIA on Different Levels
AgentFounder performs best on level 1 GAIA tasks and degrades substantially on level 3, indicating lower efficacy on more complex tasks.
- 79.5% Pass@1 and 87.2% Pass@3 are achieved on level 1 GAIA tasks.
- Figure 11 reports pass rates across different levels of the GAIA dataset.
- 50.0% Pass@1 and 58.3% Pass@3 on level 3 tasks show substantial degradation relative to level 1.The reported trend indicates that model efficacy is significantly affected by task complexity.
B.4 MoE Activations
After continual pre-training, the model’s MoE expert activations become more balanced in final layers, potentially supporting diversified expert use and training stability.
- The activation analysis displays the top eight post-softmax router scores for question-final tokens.
- CPT makes the final-layer MoE expert distribution more balanced rather than concentrated.The comparison uses router-logit activations averaged over BrowseComp-zh samples, contrasting the baseline, AgentFounder-30B-A3B, and their difference.
- Balanced expert use potentially improves post-training stability and reduces overfitting risk by mitigating dead experts.Dead experts are described as experts that cease activation and fail to contribute meaningfully to performance.
B.5 Accuracy vs. Tool Call Turns
Across several deep-research benchmarks, lower tool-call counts generally coincide with higher final accuracy, while omitting tools substantially reduces accuracy.
- The analysis measures total tool calls and task success across BrowseComp-en, BrowseComp-zh, GAIA4, and Xbench DeepResearch trajectories.Three rollouts are used to assess stability.
- Accuracy drops from 81.3% to 66.7% when no tool calls are made, although the no-tool sample is very small.The analysis therefore indicates that the model benefits from tool usage, while the small sample limits the strength of this comparison.
- Tasks with fewer tool-call turns generally achieve higher final accuracy, suggesting more efficient task completion.The authors connect this pattern to improved planning capabilities.
C Case Study
The case study identifies Paris as the setting and concludes that Riyadh Air made the balanced aircraft commitment, with firm orders matching options.
- Avelo Airlines is another balanced-order candidate, with 50 firm Embraer E2 orders and 50 additional purchase rights.
- Riyadh Air placed 25 firm orders and 25 options at the 2025 Paris Air Show.
- ASL Aviation is also associated with equal firm orders and purchase rights at the 2019 Paris Air Show, creating an alternative historical match.
- The marketplace is identified as the Paris Air Show, connected to Paris through the Louvre and Paris-related civic events.