1 00:00:03,000 --> 00:00:05,880 Evan: Welcome to Daily Paper Cast! 2 00:00:05,928 --> 00:00:14,388 Ashley: Today’s paper is from the Hugging Face daily paper list of September 29, 2026, and has received 27 upvotes. 3 00:00:14,448 --> 00:00:20,598 Evan: The title of the paper is 'CompoWorld: Compositional Environment Scaling for General Agents'. 4 00:00:20,598 --> 00:00:27,628 It's authored by Xiao-Wen Yang and Weiyi Xu, with corresponding author Wen Da from the AllSpark Team. 5 00:00:27,732 --> 00:00:30,052 Evan: Let’s dive right into the introduction. 6 00:00:30,096 --> 00:00:36,006 Ashley: Large language models, or LLMs, have evolved beyond simple text generators. 7 00:00:36,006 --> 00:00:41,976 They are now agents that can reason, utilize tools, and take actions in digital environments. 8 00:00:42,024 --> 00:00:44,594 Evan: That’s a significant leap in capability. 9 00:00:44,594 --> 00:00:48,644 But training these advanced agents comes with its own set of challenges. 10 00:00:48,696 --> 00:00:49,926 Ashley: Exactly. 11 00:00:49,926 --> 00:00:52,946 Training requires more than static demonstrations. 12 00:00:52,946 --> 00:00:59,356 You need interactive environments where actions can change external states, and outcomes can be evaluated. 13 00:00:59,400 --> 00:01:05,500 Evan: So, environment synthesis and evaluation become central to the learning process of these agents. 14 00:01:05,544 --> 00:01:08,874 Ashley: Yes, and that's where environment scaling steps in. 15 00:01:08,874 --> 00:01:15,404 It involves expanding the diversity of environments and verifiable tasks available for training these agents. 16 00:01:15,456 --> 00:01:19,176 Evan: Can you give us some context on how this has been approached so far? 17 00:01:19,224 --> 00:01:19,994 Ashley: Sure. 18 00:01:19,994 --> 00:01:28,604 Previous efforts like AgentScaler, ScaleEnv, and Agent-World have been automating the construction of tool-interaction environments and tasks. 19 00:01:28,656 --> 00:01:31,806 Evan: These efforts likely widen the scope of environments. 20 00:01:31,806 --> 00:01:34,636 But what’s the new angle this paper is exploring? 21 00:01:34,680 --> 00:01:40,680 Ashley: Alongside expanding the variety of environments, the paper targets scaling the dependencies between them. 22 00:01:40,680 --> 00:01:44,000 Real-world workflows often span multiple systems. 23 00:01:44,000 --> 00:01:50,940 For example, an agent might query a database, consult a manual, and notify stakeholders via email. 24 00:01:51,000 --> 00:01:52,220 Evan: That makes sense. 25 00:01:52,220 --> 00:01:57,420 Real-world scenarios aren’t siloed; they require coordination across systems. 26 00:01:57,480 --> 00:01:58,660 Ashley: Precisely. 27 00:01:58,660 --> 00:02:05,600 Existing benchmarks like AppWorld and Terminal-Universe already expose the need for cross-application workflows. 28 00:02:05,600 --> 00:02:11,600 But this paper introduces a new framework named Compositional Environment Scaling, or CompoWorld. 29 00:02:11,724 --> 00:02:12,944 Evan: Interesting. 30 00:02:12,944 --> 00:02:15,324 How does CompoWorld approach this scaling? 31 00:02:15,384 --> 00:02:20,964 Ashley: CompoWorld constructs tasks by composing a finite library of reusable services. 32 00:02:20,964 --> 00:02:24,894 Each service represents a typed state and provides tool sets. 33 00:02:24,894 --> 00:02:29,664 These services are then interconnected through task-specific dependency graphs. 34 00:02:29,772 --> 00:02:34,552 Evan: So, it's like creating a network of services that can be flexibly recombined? 35 00:02:34,608 --> 00:02:35,698 Ashley: Exactly. 36 00:02:35,698 --> 00:02:41,278 This setup allows the creation of a much larger space of workflows from a limited set of services. 37 00:02:41,278 --> 00:02:45,988 The agent's goal then becomes coordinating these services in novel combinations. 38 00:02:46,072 --> 00:02:46,992 Evan: I see. 39 00:02:46,992 --> 00:02:51,852 But ensuring these services work harmoniously must present some challenges. 40 00:02:51,912 --> 00:02:52,682 Ashley: Indeed. 41 00:02:52,682 --> 00:03:07,192 The authors identify three main challenges: ensuring services execute reliably, creating meaningful cross-service dependencies that are solvable, and providing useful supervision for both supervised fine-tuning and reinforcement learning. 42 00:03:07,248 --> 00:03:09,968 Evan: How does CompoWorld address these challenges? 43 00:03:10,032 --> 00:03:17,782 Ashley: To ensure reliable execution, CompoWorld standardizes service states and interfaces using typed Python schemas. 44 00:03:17,782 --> 00:03:23,592 Coding agents then build executable mock services that can be verified and repaired automatically. 45 00:03:23,700 --> 00:03:26,120 Evan: And what about the cross-service dependencies? 46 00:03:26,184 --> 00:03:32,224 Ashley: CompoWorld uses a random-walk procedure to sample services and construct dependency graphs. 47 00:03:32,224 --> 00:03:38,644 These graphs show how information flows between services without prescribing a fixed sequence of tool calls. 48 00:03:38,688 --> 00:03:40,188 Evan: That sounds robust. 49 00:03:40,188 --> 00:03:42,488 And how are these tasks used for training? 50 00:03:42,552 --> 00:03:51,672 Ashley: The framework uses verified successful trajectories for supervised fine-tuning and a Completion-Focused Rubric Reward for reinforcement learning. 51 00:03:51,672 --> 00:03:57,252 The rubric reward assigns higher weights to criteria with lower pass rates within each group. 52 00:03:57,402 --> 00:04:03,452 Evan: It seems like this approach can significantly enhance an agent's ability to manage complex workflows. 53 00:04:03,504 --> 00:04:10,034 Ashley: And CompoWorld has already constructed 448 services exposing over ten thousand tools. 54 00:04:10,034 --> 00:04:17,124 These are used to train large language models, resulting in significant performance boosts across multiple benchmarks. 55 00:04:17,184 --> 00:04:18,614 Evan: That's impressive. 56 00:04:18,614 --> 00:04:22,304 So, we’ve covered the background and objectives of CompoWorld. 57 00:04:22,304 --> 00:04:24,524 Shall we move on to the methods they used? 58 00:04:24,576 --> 00:04:27,046 Ashley: Yes, let's dive into that next. 59 00:04:27,046 --> 00:04:29,356 That’s the end of the Introduction section. 60 00:04:30,682 --> 00:04:33,702 Evan: Let's now delve into the methods laid out in the paper. 61 00:04:33,746 --> 00:04:34,736 Ashley: Certainly. 62 00:04:34,736 --> 00:04:40,886 The methods section is quite comprehensive and begins with how CompoWorld generates agent environments. 63 00:04:40,946 --> 00:04:41,486 Evan: Alright. 64 00:04:41,486 --> 00:04:45,446 How does CompoWorld approach the generation of agent environments? 65 00:04:45,506 --> 00:04:54,146 Ashley: To start, CompoWorld collects machine-readable Model Context Protocol, or MCP, specifications via web crawling. 66 00:04:54,146 --> 00:05:02,786 These specifications define the tool names, descriptions, and typed parameter schemas for each service, thereby setting up the action space. 67 00:05:02,864 --> 00:05:06,174 Evan: And what about the state space and transition functions? 68 00:05:06,248 --> 00:05:12,988 Ashley: A coding agent steps in to infer the service entities and constructs Pydantic models for the environment states. 69 00:05:12,988 --> 00:05:20,978 The service states are represented as typed records and validated by Pydantic to ensure that state transitions conform to the schema. 70 00:05:21,026 --> 00:05:24,356 Evan: So, the coding agents build and validate these models. 71 00:05:24,356 --> 00:05:25,706 What happens next? 72 00:05:25,754 --> 00:05:31,774 Ashley: Once the state model is set, the coding agent implements each tool as an operation over the current state. 73 00:05:31,774 --> 00:05:39,934 Successfully invoking a tool updates the state and provides a structured observation, all encapsulated in JSON format for uniformity. 74 00:05:39,986 --> 00:05:42,026 Evan: That seems like a solid foundation. 75 00:05:42,026 --> 00:05:45,066 But how does the system ensure these tools are reliable? 76 00:05:45,122 --> 00:05:58,482 Ashley: For reliability, the coding agent generates a corresponding test suite and executes it iteratively, repairing implementation issues until both basic successful-use cases and error-handling cases pass. 77 00:05:58,538 --> 00:06:02,278 Evan: But self-generated tests might miss certain blind spots. 78 00:06:02,278 --> 00:06:04,218 How does CompoWorld deal with this? 79 00:06:04,304 --> 00:06:05,234 Ashley: Good point. 80 00:06:05,234 --> 00:06:09,224 CompoWorld addresses this by conducting an independent validation pass. 81 00:06:09,224 --> 00:06:17,734 Another agent session generates adversarial test cases based on tool specifications to catch boundary conditions that the initial tests might have missed. 82 00:06:17,786 --> 00:06:19,526 Evan: That sounds thorough. 83 00:06:19,526 --> 00:06:23,906 What if some tools simply cannot be faithfully implemented in the local code? 84 00:06:23,954 --> 00:06:30,614 Ashley: For tools that can't be faithfully implemented, CompoWorld uses a language world model to simulate their behavior. 85 00:06:30,614 --> 00:06:36,654 The model predicts the necessary state updates and observations based on the current state and invoked action. 86 00:06:36,698 --> 00:06:37,928 Evan: Interesting. 87 00:06:37,928 --> 00:06:43,098 So, this hybrid approach ensures robust execution across all tools. 88 00:06:43,098 --> 00:06:46,118 How broad is the corpus created by this process? 89 00:06:46,178 --> 00:06:53,978 Ashley: The pipeline generates 448 independently executable services spanning 10,130 tools. 90 00:06:53,978 --> 00:07:03,358 These services are organized into broad application domains like software development, productivity, data analytics, and social media, among others. 91 00:07:03,410 --> 00:07:07,690 Evan: And these services are reusable and can be combined in various ways? 92 00:07:07,754 --> 00:07:08,804 Ashley: Exactly. 93 00:07:08,804 --> 00:07:11,804 This forms the backbone for scaling the environments. 94 00:07:11,804 --> 00:07:18,234 Composing these services allows for a vast space of training environments, supporting diverse workflows. 95 00:07:18,290 --> 00:07:25,550 Evan: Now that we understand how the environments are composed, how does CompoWorld generate specific tasks from these environments? 96 00:07:25,660 --> 00:07:35,110 Ashley: A task in CompoWorld is represented as a combination of an instruction, initial state, dependency graph, reachable goal state, and a verifier. 97 00:07:35,110 --> 00:07:40,630 The task generation starts by sampling a set of services and forming a product environment. 98 00:07:40,742 --> 00:07:43,962 Evan: And how are the dependencies between services defined? 99 00:07:44,018 --> 00:07:50,968 Ashley: Dependencies are mapped out using a random-walk procedure, which connects the services based on informational dependencies. 100 00:07:50,968 --> 00:07:58,238 This doesn’t dictate a fixed sequence of tool calls but requires the agent to navigate dependencies between services dynamically. 101 00:07:58,298 --> 00:08:02,138 Evan: Right, that requires flexible adaptation by the agent. 102 00:08:02,138 --> 00:08:05,098 What's the role of the coding agent within this setup? 103 00:08:05,162 --> 00:08:15,102 Ashley: Within the pi harness, the coding agent explores services, constructs validated initial states and instructions, and formulates a structured rubric of criteria. 104 00:08:15,102 --> 00:08:20,002 It then executes the task to generate a reference goal state through multiple phases. 105 00:08:20,066 --> 00:08:25,846 Evan: So, the task isn't just created but also validated for reachability and correctness? 106 00:08:25,898 --> 00:08:26,928 Ashley: Precisely. 107 00:08:26,928 --> 00:08:30,898 This probe defines both the goal state and the criteria for success. 108 00:08:30,898 --> 00:08:36,438 When the reference goal is reachable and the rubric accepts the result, the task is considered valid. 109 00:08:36,482 --> 00:08:38,972 Evan: I understand the task generation now. 110 00:08:38,972 --> 00:08:42,022 How does CompoWorld use these tasks for training? 111 00:08:42,074 --> 00:08:48,644 Ashley: The verified tasks are initially used for supervised fine-tuning, establishing a strong base policy. 112 00:08:48,644 --> 00:08:56,994 Then, reinforcement learning takes over, using rubric rewards that focus on unsatisfied criteria to drive full task completion. 113 00:08:57,050 --> 00:09:01,470 Evan: And the Completion-Focused Rubric Reward plays a crucial role here? 114 00:09:01,514 --> 00:09:02,304 Ashley: Yes. 115 00:09:02,304 --> 00:09:10,244 This reward mechanism assigns greater weight to criteria that are less frequently met, encouraging the agent to focus on those areas. 116 00:09:10,244 --> 00:09:16,874 It's combined with Group Relative Policy Optimization, which normalizes rewards within each rollout group. 117 00:09:16,982 --> 00:09:17,812 Evan: I see. 118 00:09:17,812 --> 00:09:21,822 This fine-tuning and reinforcement combination sounds impactful. 119 00:09:21,822 --> 00:09:24,462 How has this approach performed experimentally? 120 00:09:24,506 --> 00:09:36,176 Ashley: Experiments included training Qwen3.6-35B-A3B on 3,000 supervised fine-tuning trajectories followed by 1,000 reinforcement learning tasks. 121 00:09:36,176 --> 00:09:42,366 The results show an average gain of 9.17 points over the baseline across eight benchmarks. 122 00:09:42,410 --> 00:09:44,630 Evan: That's a substantial improvement. 123 00:09:44,630 --> 00:09:47,390 Any particular benchmarks where it stood out? 124 00:09:47,450 --> 00:09:57,240 Ashley: On AutomationBench, CompoWorld tripled the baseline's pass rate and even surpassed models like GPT-5.4 and Claude Opus 4.6. 125 00:09:57,240 --> 00:10:01,830 It also led all six compared agent-specialized models on this benchmark. 126 00:10:01,874 --> 00:10:04,514 Evan: Wow, that’s impressive. 127 00:10:04,562 --> 00:10:05,462 Ashley: Indeed. 128 00:10:05,462 --> 00:10:15,482 The method section of this paper shows how compositional environment scaling can significantly enhance agent training, offering new potential for complex workflow management. 129 00:10:15,530 --> 00:10:17,130 Evan: Great insights, Ashley. 130 00:10:17,130 --> 00:10:19,630 This brings us to the end of the Methods section. 131 00:10:20,923 --> 00:10:25,603 Evan: Okay, let's move on to the experiments and results presented in the paper. 132 00:10:25,659 --> 00:10:36,319 Ashley: In the experiments section, the authors aim to evaluate how training in composed environments impacts agent performance across different domains, comparing against a range of models. 133 00:10:36,363 --> 00:10:37,673 Evan: That makes sense. 134 00:10:37,673 --> 00:10:42,453 Establishing how well CompoWorld performs compared to other models is crucial. 135 00:10:42,453 --> 00:10:45,103 What are the baselines they used in the comparison? 136 00:10:45,147 --> 00:11:04,547 Ashley: The paper compares CompoWorld with several frontier models, including closed-source models like GPT-5.4, Claude Opus 4.6, and Gemini-3.1 Pro, as well as open-weight models like DeepSeek-V4-Flash, GLM-5.2, and various versions of Qwen. 137 00:11:04,611 --> 00:11:08,291 Evan: And there are benchmarks for agent-specialized models as well? 138 00:11:08,355 --> 00:11:09,315 Ashley: Correct. 139 00:11:09,315 --> 00:11:26,515 They also compared against agent-specialized models at the same scale, including Apodex 1.1 Mini, Occamy-1.0, Agents-A1, Nex-N2-mini, BigBang-1.0, and Ornith-1.5-35B. 140 00:11:26,631 --> 00:11:29,021 Evan: A comprehensive set of comparisons. 141 00:11:29,021 --> 00:11:31,871 What benchmarks did they use for the evaluation? 142 00:11:31,923 --> 00:11:59,863 Ashley: Evaluation covered eight challenging benchmarks, such as τ-Knowledge Banking for customer support, WildClawBench for real-world workflows, SkillsBench for reusable skills, AutomationBench for cross-application business workflows, DeepPlanning for planning under verifiable constraints, VitaBench and VitaBench 2.0 for multi-turn interactions and long-term user assistance, and ALE, the Agents’ Last Exam, for professional task performance. 143 00:11:59,907 --> 00:12:01,397 Evan: That's quite diverse. 144 00:12:01,397 --> 00:12:04,487 How did CompoWorld perform across these benchmarks? 145 00:12:04,539 --> 00:12:15,849 Ashley: CompoWorld showed significant improvements across all eight benchmarks, averaging a 9.17-point gain over the baseline model Qwen3.6-35B-A3B. 146 00:12:15,849 --> 00:12:30,319 On AutomationBench, it tripled the pass rate from 10.33% to 32.33%, outperforming models like GPT-5.4, Gemini-3.1 Pro, and nearly matching DeepSeek-V4-Flash. 147 00:12:30,423 --> 00:12:33,313 Evan: Tripling the pass rate is remarkable! 148 00:12:33,313 --> 00:12:36,003 Were there any other notable performances? 149 00:12:36,051 --> 00:12:41,391 Ashley: For instance, on SkillsBench, CompoWorld improved by 15.19 points. 150 00:12:41,391 --> 00:12:47,781 VitaBench saw a gain of 10.50 points, indicating strong results in multi-turn service interactions. 151 00:12:47,781 --> 00:12:55,511 However, the gain on VitaBench 2.0 was smaller, suggesting more limited transfer to personalized, long-term assistance. 152 00:12:55,563 --> 00:12:59,803 Evan: What about the impact on partial credit scores within different domains? 153 00:12:59,859 --> 00:13:03,799 Ashley: Domains like HR and Marketing showed the largest gains. 154 00:13:03,799 --> 00:13:11,479 CompoWorld raised the average domain score from 41.94 to 72.68 in AutomationBench. 155 00:13:11,479 --> 00:13:18,979 HR showed an increase of 49.40 points, while Marketing improved by 32.51 points. 156 00:13:19,035 --> 00:13:23,855 Evan: It’s great to see such widespread improvement across various business functions. 157 00:13:23,855 --> 00:13:26,375 How did these gains compare with other models? 158 00:13:26,427 --> 00:13:33,957 Ashley: CompoWorld outperformed both GPT-5.4 and GLM-5.2 in five of six domains. 159 00:13:33,957 --> 00:13:39,627 It ranked first in HR and second in Finance and Operations among the listed models. 160 00:13:39,675 --> 00:13:40,905 Evan: That's impressive. 161 00:13:40,905 --> 00:13:47,815 So, CompoWorld definitely shows strong performance in both specific benchmarks and broader domain applications. 162 00:13:47,859 --> 00:13:48,749 Ashley: Indeed. 163 00:13:48,749 --> 00:13:55,079 The evaluation also looked at how agent performance changes as the number of training environments scales. 164 00:13:55,079 --> 00:14:04,059 CompoWorld allows expansion of training environments without needing new services for each example, using a compositional approach with reusable services. 165 00:14:04,187 --> 00:14:07,727 Evan: How does the increase in training environments affect performance? 166 00:14:07,779 --> 00:14:15,499 Ashley: Interestingly, all four benchmarks showed improvement over the baseline with as few as 100 supervised fine-tuning examples. 167 00:14:15,499 --> 00:14:28,799 At 3,000 examples, τ-Knowledge Banking rose from 10.65 points to 17.87 points, and DeepPlanning saw a jump from 26.04 points to 35.02 points. 168 00:14:28,911 --> 00:14:32,441 Evan: It demonstrates the strong impact of environment scaling. 169 00:14:32,441 --> 00:14:36,091 What are the key takeaways here regarding environment composition? 170 00:14:36,147 --> 00:14:43,067 Ashley: The findings suggest that composed-environment training yields substantial performance improvements over single-environment training. 171 00:14:43,067 --> 00:14:51,867 For example, τ-Knowledge Banking improved from 9.62 to 17.53 points when using composed environments. 172 00:14:51,915 --> 00:14:56,235 Evan: So, composing environments really boosts the training outcome. 173 00:14:56,235 --> 00:14:59,535 What insights do we have from reinforcement learning dynamics? 174 00:14:59,625 --> 00:15:06,535 Ashley: During reinforcement learning, the completion-focused rubric rewards aimed to allocate learning to unmet criteria. 175 00:15:06,535 --> 00:15:15,695 The learning curve showed initial adjustment followed by sustained progress, reaching about 0.79 in mean trajectory reward by step 150. 176 00:15:15,747 --> 00:15:19,487 Evan: How did the use of reweighted rewards impact the training? 177 00:15:19,539 --> 00:15:25,659 Ashley: Using reweighted rewards to focus on less frequently met criteria showed more stable gains. 178 00:15:25,659 --> 00:15:37,599 Evaluation-set scores with reweighting moved ahead by step 50 and surpassed scores without reweighting, creating a gap of about 0.09 points by step 140. 179 00:15:37,659 --> 00:15:41,229 Evan: That highlights the effectiveness of the completion-focused rubric. 180 00:15:41,229 --> 00:15:44,019 What’s the overall conclusion from these experiments? 181 00:15:44,067 --> 00:15:55,087 Ashley: Overall, the results demonstrate that CompoWorld’s approach to compositional environment scaling provides better training signals and robust task management across complex workflows. 182 00:15:55,087 --> 00:16:02,187 It highlights how scaling training environments along compositional lines can significantly enhance agent performance. 183 00:16:02,235 --> 00:16:06,035 Evan: That wraps up our discussion of the experiments and results section. 184 00:16:07,301 --> 00:16:11,801 Evan: Moving on, let's explore the Related Work section discussed in the paper. 185 00:16:11,861 --> 00:16:20,971 Ashley: The Related Work section covers existing research on environment scaling for large language models and compositional generalization for agent learning. 186 00:16:20,971 --> 00:16:25,201 These two areas provide the foundational context for CompoWorld. 187 00:16:25,253 --> 00:16:27,493 Evan: Sounds like a comprehensive background. 188 00:16:27,493 --> 00:16:29,403 Let's start with environment scaling. 189 00:16:29,403 --> 00:16:30,973 How has this been approached? 190 00:16:31,037 --> 00:16:38,417 Ashley: Automated environment construction has been pivotal in reducing reliance on real services and supporting interactive agent training. 191 00:16:38,417 --> 00:16:45,077 For example, DreamGym utilizes reasoning-based experience models to simulate transitions and feedback. 192 00:16:45,155 --> 00:16:47,085 Evan: DreamGym sounds interesting. 193 00:16:47,085 --> 00:16:48,905 Are there other notable methods? 194 00:16:48,965 --> 00:16:50,565 Ashley: Yes, certainly. 195 00:16:50,565 --> 00:17:00,815 Executable methods for environment synthesis and verifiable tasks, showcased in studies like AutoForge, provide structured environments for agentic reinforcement learning. 196 00:17:00,815 --> 00:17:07,605 EnvScaler separates environment skeleton construction from scenario generation and rule-based validation. 197 00:17:07,661 --> 00:17:12,281 Evan: So, different approaches have been employed to create interactive environments. 198 00:17:12,281 --> 00:17:14,781 How about dependency-graph expansion? 199 00:17:14,867 --> 00:17:25,407 Ashley: Dependency-graph expansion and topology-aware trajectory synthesis, as seen in works like ScaleEnv, allow for the creation of complex interaction scenarios. 200 00:17:25,407 --> 00:17:30,397 Similarly, AutoWebWorld generates interactive websites for agent training. 201 00:17:30,461 --> 00:17:33,141 Evan: What about configuring existing software? 202 00:17:33,141 --> 00:17:34,941 Is that something explored too? 203 00:17:35,787 --> 00:17:44,867 Ashley: Gym-Anything transforms existing software applications into agent environments, enabling practical agent training based on real software interactions. 204 00:17:44,867 --> 00:17:49,737 This approach provides realistic data and supports a wide range of tasks. 205 00:17:49,781 --> 00:17:53,881 Evan: Sounds like there’s a host of strategies to simulate environments for training. 206 00:17:53,881 --> 00:17:56,421 How does CompoWorld build on these methods? 207 00:17:56,477 --> 00:18:04,327 Ashley: CompoWorld leverages independently executable, stateful services combined into task-specific causal dependency graphs. 208 00:18:04,327 --> 00:18:12,537 This approach makes service combinations and cross-service requirements explicit, adding controllable dimensions to training-environment generation. 209 00:18:12,581 --> 00:18:13,611 Evan: Interesting. 210 00:18:13,611 --> 00:18:17,301 Now, let’s shift our focus to compositional generalization. 211 00:18:17,301 --> 00:18:18,941 What's covered in this area? 212 00:18:18,989 --> 00:18:28,969 Ashley: Compositional generalization involves reusing familiar tools and skills in new task structures, respecting dependencies between actions and environment states. 213 00:18:28,969 --> 00:18:35,689 Studies like CompWoB highlight the challenge of transferring performance from individual tasks to their combinations. 214 00:18:35,741 --> 00:18:37,951 Evan: Transferring skills seems critical. 215 00:18:37,951 --> 00:18:40,961 Are there benchmarks specifically focused on this? 216 00:18:41,021 --> 00:18:42,141 Ashley: Indeed. 217 00:18:42,141 --> 00:18:50,491 AppWorld extends evaluation to coordinate workflows across multiple applications, requiring agents to manage APIs and state changes. 218 00:18:50,491 --> 00:18:55,361 This benchmark isolates the need for handling complex interconnected tasks. 219 00:18:55,501 --> 00:19:00,001 Evan: And on the methodological front, how is compositionality tackled? 220 00:19:00,083 --> 00:19:12,483 Ashley: Voyager proposes building a library of reusable skills, while Compositional Skill Routing decomposes requests and assembles plans that respect dependencies among model context protocol skills. 221 00:19:12,483 --> 00:19:17,373 Both offer ways to equip agents for new tasks using established capabilities. 222 00:19:17,429 --> 00:19:22,049 Evan: So, combining skills and managing dependencies is a recurrent theme. 223 00:19:22,049 --> 00:19:24,509 How does CompoWorld fit into this picture? 224 00:19:24,557 --> 00:19:32,097 Ashley: CompoWorld advances this direction by recombining independently executable services along with their causal dependencies. 225 00:19:32,097 --> 00:19:41,617 The aim is to train agents to transfer their knowledge of individual services to new workflows by learning how information and state changes connect across services. 226 00:19:41,669 --> 00:19:46,679 Evan: Right, the focus on recombination of services and dependencies seems unique. 227 00:19:46,679 --> 00:19:48,309 Anything else worth mentioning? 228 00:19:48,395 --> 00:19:49,635 Ashley: Definitely. 229 00:19:49,635 --> 00:19:57,535 Terminal-Universe reconstructs workspaces from agent trajectories, creating tasks that span writable and read-only codebases. 230 00:19:57,535 --> 00:20:04,985 This framework complements CompoWorld’s idea of cross-environment workflows by focusing on interconnected workspace states. 231 00:20:05,105 --> 00:20:09,595 Evan: It’s enlightening to see how different studies intersect and build upon each other. 232 00:20:09,595 --> 00:20:11,625 Any final thoughts from this section? 233 00:20:11,669 --> 00:20:23,699 Ashley: Overall, CompoWorld leverages past research on automated environment construction and compositional generalization, pushing the boundaries of agent training through scalable, cross-service environments. 234 00:20:23,699 --> 00:20:29,189 It synthesizes these approaches to enhance workflow diversity and task complexity in agent training. 235 00:20:29,237 --> 00:20:31,587 Evan: Thanks for breaking down that section, Ashley. 236 00:20:31,587 --> 00:20:33,817 That’s the end of the Related Work section. 237 00:20:35,070 --> 00:20:37,850 Evan: We're nearing the end of our discussion on CompoWorld. 238 00:20:37,850 --> 00:20:41,190 Let's summarize the key contributions and takeaways. 239 00:20:41,238 --> 00:20:46,098 Ashley: CompoWorld makes several key contributions to the field of AI and agent training. 240 00:20:46,098 --> 00:20:56,498 First, it introduces Compositional Environment Scaling, which allows for the creation of complex and diverse training environments from a finite set of reusable services. 241 00:20:56,550 --> 00:20:58,700 Evan: That's a significant step forward. 242 00:20:58,700 --> 00:21:07,610 The ability to compose environments dynamically means we can train agents on a much larger variety of tasks without manually creating each one. 243 00:21:07,662 --> 00:21:08,692 Ashley: Exactly. 244 00:21:08,692 --> 00:21:18,582 By leveraging dependency graphs, CompoWorld facilitates meaningful information flow between different services without locking into a fixed sequence of tool calls. 245 00:21:18,582 --> 00:21:22,022 This setup mimics real-world workflows more closely. 246 00:21:22,086 --> 00:21:27,526 Evan: The paper also addresses the challenge of ensuring these services execute reliably, correct? 247 00:21:27,582 --> 00:21:29,202 Ashley: Yes, that's right, Evan. 248 00:21:29,202 --> 00:21:34,112 CompoWorld standardizes interaction interfaces using typed Python schemas. 249 00:21:34,112 --> 00:21:41,722 This allows coding agents to generate executable mock services and independent validation to catch any potential errors. 250 00:21:41,766 --> 00:21:51,086 Evan: And using a hybrid approach with a language world model ensures the tools that can’t be locally implemented still function accurately within the environment. 251 00:21:51,190 --> 00:22:00,940 Ashley: Additionally, the Completion-Focused Rubric Reward in reinforcement learning aims to drive full task completion by focusing agents on unmet criteria. 252 00:22:00,940 --> 00:22:03,730 This leads to more robust learning outcomes. 253 00:22:03,774 --> 00:22:11,434 Evan: The experimental results certainly highlight the effectiveness of this approach, showing significant improvements across multiple benchmarks. 254 00:22:11,478 --> 00:22:18,438 Ashley: In summary, CompoWorld presents a valuable framework for advancing agent training through compositional environment scaling. 255 00:22:18,438 --> 00:22:25,078 It enhances training efficiency and enables agents to handle complex, cross-domain workflows more effectively. 256 00:22:25,134 --> 00:22:27,794 Evan: That brings us to the end of today’s episode. 257 00:22:27,846 --> 00:22:30,856 Ashley: Thank you for tuning in to Daily Paper Cast. 258 00:22:30,856 --> 00:22:34,566 We hope you found our discussion on CompoWorld insightful. 259 00:22:34,614 --> 00:22:41,494 Evan: Make sure to join us next time as we continue to explore the latest advancements in AI and machine learning research. 260 00:22:41,610 --> 00:22:44,730 Ashley: Until then, stay curious and keep learning. 261 00:22:44,850 --> 00:22:46,270 Evan: Goodbye, everyone!