رونوشت 🔗
1
00:00:03,000 --> 00:00:08,580
Evan: Welcome to Daily Paper Cast, where we break down the latest in AI research.
2
00:00:08,640 --> 00:00:16,760
Ashley: Today’s paper is from the Hugging Face daily paper list of October 14, 2026, and has received 34 upvotes.
3
00:00:16,824 --> 00:00:22,144
Evan: The paper is titled 'Harness2: Recursive Agent Harnessing for an Open World'.
4
00:00:22,200 --> 00:00:29,740
Ashley: It is authored by Ruiyao Xu from Northwestern University and Yanfei Chen from Google Cloud AI Research.
5
00:00:29,740 --> 00:00:37,120
Ruiyao Xu is the corresponding author, and this work was done while Ruiyao interned at Google Cloud AI Research.
6
00:00:37,236 --> 00:00:43,876
Evan: LLM agents’ capabilities rely heavily on their execution harnesses as much as their underlying models.
7
00:00:43,876 --> 00:00:51,036
The harness includes prompts, memory, skills, tools, and orchestration that enable the model to perform tasks.
8
00:00:51,036 --> 00:00:59,016
Such an approach is critical as AI agents expand into diverse and high-stakes domains, including legal work and office workflows.
9
00:00:59,064 --> 00:01:07,264
Ashley: Recursive self-improvement, or RSI, is a method that offers a promising path to autonomously evolve these harnesses over time.
10
00:01:07,264 --> 00:01:15,684
RSI allows an AI system to make persistent improvements to its own capabilities, which then further shape subsequent iterations of improvements.
11
00:01:15,684 --> 00:01:20,984
However, deploying RSI in open-world settings exposes some key challenges.
12
00:01:21,048 --> 00:01:31,928
Evan: First, open-world tasks are often unseen during the harness design phase, and may be novel at test time, making ground-truth verifiers unreliable for guiding harness edits online.
13
00:01:31,928 --> 00:01:38,728
Second, the RSI process remains highly task-specific and struggles to generalize to novel scenarios.
14
00:01:38,784 --> 00:01:46,944
Ashley: To tackle these issues, the paper introduces Harness2, a plug-and-play framework for recursive agent harnessing in open-world environments.
15
00:01:46,944 --> 00:01:49,654
Harness2 employs two levels of recursion.
16
00:01:49,654 --> 00:02:00,084
Task-level recursion refines the current task’s harness using a contrastive Propose–Probe–Compose step, guiding edits effectively without relying on ground-truth verifiers.
17
00:02:00,084 --> 00:02:05,424
This task-level recursion can be scaled further through parallel or sequential recursion.
18
00:02:05,472 --> 00:02:12,642
Evan: Then there's harnessing-level recursion, which extracts experience from these edits and updates a continual improvement harness.
19
00:02:12,642 --> 00:02:18,832
This enables the agent to better adapt to future tasks by using lessons learned from past harness improvements.
20
00:02:18,832 --> 00:02:27,652
The combined approach aims to refine the harness for the current task while simultaneously improving the mechanisms for future harnessing across a stream of tasks.
21
00:02:27,696 --> 00:02:33,546
Ashley: This method was evaluated across three professional benchmarks and four model-harness configurations.
22
00:02:33,546 --> 00:02:50,236
Initial results show that Harness2 consistently improved performance over base harnesses, demonstrating gains of up to 11.5 points on the WorkBuddy-Bench benchmark, 4.5 points on JobBench, and 11.8 percentage points in LAB’s Partial pass rate.
23
00:02:50,280 --> 00:03:00,900
Evan: The improvements are particularly notable with increased task-level recursion budgets, indicating scalable benefits from both parallel and sequential recursion strategies.
24
00:03:00,960 --> 00:03:04,500
Ashley: That brings us to the end of the Introduction section of the paper.
25
00:03:04,500 --> 00:03:09,100
We’ll be diving deeper into the methods and experiments in the following segments.
26
00:03:10,346 --> 00:03:13,726
Evan: Let's dive into the methods described in the paper.
27
00:03:13,726 --> 00:03:17,886
Ashley, can you walk us through the proposed framework of Harness2?
28
00:03:17,930 --> 00:03:19,060
Ashley: Sure, Evan.
29
00:03:19,060 --> 00:03:25,030
Harness2 is designed to tackle the challenges of recursive agent harnessing in open-world environments.
30
00:03:25,030 --> 00:03:32,090
The framework emphasizes two interconnected levels of recursion: task-level recursion and harnessing-level recursion.
31
00:03:32,198 --> 00:03:35,368
Evan: Alright, let's start with task-level recursion.
32
00:03:35,368 --> 00:03:36,698
How does that work?
33
00:03:36,746 --> 00:03:42,086
Ashley: Task-level recursion is aimed at refining the execution harness for the current task.
34
00:03:42,086 --> 00:03:53,206
Since open-world tasks typically lack ground-truth labels or reliable verifiers, the paper introduces a contrastive improvement step known as Propose–Probe–Compose.
35
00:03:53,258 --> 00:03:58,098
Evan: Could you elaborate on the Propose–Probe–Compose step?
36
00:03:58,154 --> 00:04:03,914
Ashley: The pipeline first obtains a reference execution, which reveals how the current harness performs.
37
00:04:03,914 --> 00:04:16,964
Based on this, it proposes two harness candidates: an invariant candidate targeting agentic deficiencies like planning or tool use, and a domain-specific candidate targeting the knowledge and constraints of the current environment.
38
00:04:16,964 --> 00:04:23,114
These candidates are then executed on the same task to compare their behavior against the reference execution.
39
00:04:23,222 --> 00:04:27,482
Evan: And what does the Propose–Probe–Compose step achieve?
40
00:04:27,530 --> 00:04:34,220
Ashley: The behavioral contrasts obtained from probing the candidate harnesses guide the composition of a successor harness.
41
00:04:34,220 --> 00:04:48,430
This composition selectively merges supported changes, revises conflicting edits, adopts one candidate, or preserves the current harness, ensuring each decision is based on concrete observations rather
Load full transcript
than presumed score gains.
42
00:04:48,482 --> 00:04:51,122
Evan: That sounds like a meticulous process.
43
00:04:51,122 --> 00:04:54,762
How does parallel and sequential recursion fit into this?
44
00:04:54,818 --> 00:05:01,148
Ashley: Parallel recursion involves proposing multiple candidates and composing refinements from their execution evidence.
45
00:05:01,148 --> 00:05:06,968
This process runs concurrently, aggregating evidence from various sources to improve the harness.
46
00:05:06,968 --> 00:05:11,178
On the other hand, sequential recursion applies iterative self-refinement.
47
00:05:11,178 --> 00:05:19,838
It carries the refined harness and accumulated execution evidence into successive steps, building on each prior step to further refine the harness.
48
00:05:19,958 --> 00:05:22,668
Evan: So that's how task-level recursion works.
49
00:05:22,668 --> 00:05:24,878
What about harnessing-level recursion?
50
00:05:24,938 --> 00:05:32,808
Ashley: Harnessing-level recursion focuses on using the experience from task-level explorations to update a continual improvement harness.
51
00:05:32,808 --> 00:05:40,688
After each task, the improvement process is reviewed, extracting lessons about effective changes and the conditions under which they worked.
52
00:05:40,688 --> 00:05:47,138
These lessons are retained in a harness improvement library, which is then used to guide future harnessing efforts.
53
00:05:47,186 --> 00:05:48,406
Evan: Interesting.
54
00:05:48,406 --> 00:05:51,766
How does the harness improvement library work exactly?
55
00:05:51,818 --> 00:06:02,968
Ashley: The improvement library stores adopted harnesses along with detailed EXPERIENCE.md files that record applicable situations, lessons learned, and supporting evidence.
56
00:06:02,968 --> 00:06:10,988
This ensures that each reusable component is documented, enabling it to be inspected and adapted by subsequent tasks.
57
00:06:10,988 --> 00:06:19,938
The library helps accumulate valuable execution and composition decisions, benefiting future harnessing efforts without requiring additional probing.
58
00:06:20,054 --> 00:06:26,134
Evan: What happens if there isn't sufficient evidence to support new entries in the improvement library?
59
00:06:26,186 --> 00:06:31,506
Ashley: In cases where evidence is insufficient, no new experience entry is added to the library.
60
00:06:31,506 --> 00:06:39,316
The focus remains on retaining executable harnesses and supporting evidence that can reliably guide future proposals and compositions.
61
00:06:39,316 --> 00:06:44,706
This selective approach ensures that only well-supported lessons influence future harnessing.
62
00:06:44,762 --> 00:06:50,042
Evan: What kind of benchmarks and configurations did the paper evaluate Harness2 on?
63
00:06:50,090 --> 00:06:58,060
Ashley: The framework was evaluated on WorkBuddy-Bench, JobBench, and LAB, across four executor-harness configurations.
64
00:06:58,060 --> 00:07:06,290
These benchmarks cover distinct
professional domains such as office-style deliverables, end-to-end professional job tasks, and legal work.
65
00:07:06,290 --> 00:07:13,010
The study aimed to demonstrate consistent gains in performance by comparing Harness2 against several baselines.
66
00:07:13,118 --> 00:07:16,698
Evan: And how did Harness2 perform on these benchmarks?
67
00:07:16,754 --> 00:07:30,864
Ashley: Harness2 showed improved performance in every configuration, with gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points in LAB’s Partial pass rate.
68
00:07:30,864 --> 00:07:40,754
The improvements were particularly notable with increased task-level recursion budgets, indicating scalable benefits from both parallel and sequential recursion strategies.
69
00:07:40,802 --> 00:07:42,402
Evan: That's quite impressive.
70
00:07:42,402 --> 00:07:47,722
Were there any insights regarding variability among different models or configurations?
71
00:07:47,786 --> 00:07:59,236
Ashley: Indeed, weaker executors like GLM-5 benefited more from harness improvements on WorkBuddy-Bench and LAB compared to stronger models like Gemini 3.5 Flash.
72
00:07:59,236 --> 00:08:09,426
This suggests that harness refinement can help weaker models outperform stronger ones under certain conditions, highlighting the significant impact of well-crafted harnesses.
73
00:08:09,482 --> 00:08:15,642
Evan: With such promising results, what are the potential limitations of Harness2 mentioned in the paper?
74
00:08:15,698 --> 00:08:21,708
Ashley: While Harness2 delivers consistent improvements, there are costs associated with recursive refinement.
75
00:08:21,708 --> 00:08:26,758
Employing additional executions and editor sessions increases computation demands.
76
00:08:26,758 --> 00:08:36,978
Moreover, the method relies on behavioral contrasts and model judgment without ground-truth supervision, which may not guarantee correctness or ease of transfer between tasks.
77
00:08:37,114 --> 00:08:41,334
Evan: And what future directions does the paper propose for Harness2?
78
00:08:41,378 --> 00:08:52,148
Ashley: The paper suggests managing the improvement harness to consolidate overlapping components, revise outdated experience, and refine retrieval according to the current task and evidence.
79
00:08:52,148 --> 00:09:04,358
Moreover, adaptive recursion is proposed to dynamically allocate exploration based on task difficulty and relevance of prior experience, potentially balancing the cost of refinement with its benefits.
80
00:09:04,418 --> 00:09:05,468
Evan: Fascinating.
81
00:09:05,468 --> 00:09:07,398
That's the end of the Method section.
82
00:09:07,398 --> 00:09:11,558
Next, we'll look into the experiments and results presented in the paper.
83
00:09:12,819 --> 00:09:18,809
Evan: Alright Ashley, let’s delve into the experimental setup and results that this paper presents.
84
00:09:18,809 --> 00:09:23,439
What benchmarks were used to evaluate the performance of Harness2?
85
00:09:23,499 --> 00:09:33,009
Ashley: The authors evaluated Harness2 on three benchmarks: WorkBuddy-Bench, JobBench, and Harvey’s Legal Agent Benchmark, referred to as LAB.
86
00:09:33,009 --> 00:09:42,419
These benchmarks span distinct professional domains including office-style tasks in multiple categories, end-to-end professional job tasks, and legal work.
87
00:09:42,543 --> 00:09:45,983
Evan: And how were the benchmarks structured for the evaluation?
88
00:09:46,065 --> 00:09:51,715
Ashley: WorkBuddy-Bench includes tasks from three different domains: Code, Web, and Office.
89
00:09:51,715 --> 00:09:57,395
JobBench targets long, end-to-end professional tasks across various occupational domains.
90
00:09:57,395 --> 00:10:04,215
LAB focuses entirely on legal tasks, evaluating criteria-specific pass rates without executable checks.
91
00:10:04,215 --> 00:10:12,135
Tasks within these benchmarks are processed individually, and the harness improvement library starts empty but updates after each task.
92
00:10:12,195 --> 00:10:16,335
Evan: What models and configurations were tested with Harness2?
93
00:10:16,395 --> 00:10:28,115
Ashley: The experiments involved four executor–harness configurations utilizing Gemini 3.5 Flash and GLM-5 models, under both OpenCode and Codex initial harnesses.
94
00:10:28,115 --> 00:10:34,975
Deploying Harness2 in diverse configurations aimed to demonstrate consistent improvements across different setups.
95
00:10:35,019 --> 00:10:38,139
Evan: That's a comprehensive evaluation matrix.
96
00:10:38,139 --> 00:10:40,859
What were the primary results for Harness2?
97
00:10:40,923 --> 00:10:45,823
Ashley: Harness2 delivered consistent gains across all configurations and benchmarks.
98
00:10:45,823 --> 00:10:54,623
For example, a single improvement step improved the GLM-5 average score by 11.5 points on WorkBuddy-Bench with the OpenCode harness.
99
00:10:54,623 --> 00:11:05,303
The framework raised scores across three domains, including maximum increases of 4.5 points on JobBench and 11.8 percentage points in LAB’s Partial pass rate.
100
00:11:05,425 --> 00:11:08,995
Evan: How did task-level recursion impact the performance?
101
00:11:09,051 --> 00:11:14,241
Ashley: Increasing the task-level recursion budget demonstrated notable performance enhancements.
102
00:11:14,241 --> 00:11:31,401
Parallel recursion at a scaling budget of five added further improvements, approximately 2.9 to 3.9 points on WorkBuddy-Bench, 3.0 to 4.2 points on JobBench, and 2.6 to 6.7 percentage points in LAB’s Partial pass rate.
103
00:11:31,401 --> 00:11:40,331
The study also indicated that weaker executors benefited more from harness refinement, often outperforming stronger models under similar conditions.
104
00:11:40,425 --> 00:11:44,955
Evan: What insights were gained regarding the compute–performance trade-off?
105
00:11:45,003 --> 00:11:52,723
Ashley: Harness2's configurations lie on the empirical Pareto frontier, balancing
performance gains with total computational costs.
106
00:11:52,723 --> 00:12:00,893
For instance, raising the recursion budget from 1 to 5 saw compute costs increase, but also resulted in higher performance gains.
107
00:12:00,893 --> 00:12:07,683
The study found that ReasoningBank and Harness Scaling yielded smaller gains while consuming significant compute resources.
108
00:12:07,683 --> 00:12:12,963
This emphasizes the importance of efficient harness refinement in maximizing performance.
109
00:12:13,071 --> 00:12:16,241
Evan: And what about the reuse of harness improvements?
110
00:12:16,241 --> 00:12:17,891
Did the study discuss this?
111
00:12:17,955 --> 00:12:25,995
Ashley: Yes, the study analyzed how well reusable components and experience from past tasks contributed to future harnessing efforts.
112
00:12:25,995 --> 00:12:34,415
Library entries include executable harnesses and EXPERIENCE.md files, capturing applicable situations and supporting evidence.
113
00:12:34,415 --> 00:12:42,795
This systematic retention and reuse of improvements helped avoid redundant probing, demonstrating gains without excessive recomputation.
114
00:12:42,843 --> 00:12:48,343
Evan: Were there any specific comparisons made between Harness2 and production harnesses?
115
00:12:48,387 --> 00:12:55,267
Ashley: Harness2 was compared with fixed production harnesses like CodeBuddy, OpenCode, Codex, and Claude Code.
116
00:12:55,267 --> 00:13:01,137
In every domain, Harness2 starting from OpenCode outperformed the strongest fixed alternative.
117
00:13:01,137 --> 00:13:09,707
The incremental gains ranged between 4.6 to 9.4 points, proving the advantage of refining existing harnesses dynamically.
118
00:13:09,871 --> 00:13:12,591
Evan: How about tool-use reliability?
119
00:13:12,591 --> 00:13:15,411
What did the paper find regarding this aspect?
120
00:13:15,459 --> 00:13:24,479
Ashley: Harness2 lowered tool-error rates across all benchmarks, making fewer calls on WorkBuddy-Bench and more on JobBench and LAB.
121
00:13:24,479 --> 00:13:32,579
Improved reliability enhanced performance, highlighting efficient tool usage as a core benefit of recursive harnessing strategies.
122
00:13:32,643 --> 00:13:37,583
Evan: Were there any notable cases where Harness2 did not perform as expected?
123
00:13:37,635 --> 00:13:41,595
Ashley: The paper identified some failure cases under specific conditions.
124
00:13:41,595 --> 00:13:51,995
Issues like blind selection of weaker candidates, incorrect premises propagating across candidates, and variability in execution outcomes challenged consistent performance.
125
00:13:51,995 --> 00:13:58,835
These cases underscore the need for balanced coverage and robust probing mechanisms to mitigate task complexities.
126
00:13:58,899 --> 00:14:04,679
Evan: Finally, what contribution does each harness component have towards improving performance?
127
00:14:04,761 --> 00:14:07,761
Ashley: All components contributed to varying degrees.
128
00:14:07,761 --> 00:14:16,321
Removing prompt-related components, skills, or scripts from an improved harness demonstrated measurable losses, reinforcing their importance.
129
00:14:16,321 --> 00:14:27,311
Skills and scripts accounted for roughly half the edits in each benchmark, and systematic retention of reusable elements in the improvement library helped accumulate valuable harness improvements.
130
00:14:27,363 --> 00:14:28,943
Evan: Interesting work indeed.
131
00:14:28,943 --> 00:14:32,623
So, that concludes the Experiment section of the paper.
132
00:14:33,929 --> 00:14:37,459
Evan: Let’s move on to the related work discussed in the paper.
133
00:14:37,459 --> 00:14:39,889
Ashley, can you give us an overview?
134
00:14:39,971 --> 00:14:41,161
Ashley: Of course, Evan.
135
00:14:41,161 --> 00:14:50,321
The related work section in the paper is quite extensive and dives deep into the various facets of recursive self-improvement and harness optimization.
136
00:14:50,381 --> 00:14:52,061
Evan: That sounds interesting.
137
00:14:52,061 --> 00:14:56,941
What exactly is recursive self-improvement in the context of AI?
138
00:14:57,005 --> 00:15:07,775
Ashley: Recursive self-improvement, or RSI, is a concept where AI systems participate in their own development by making persistent improvements to their capabilities.
139
00:15:07,775 --> 00:15:13,795
These improvements then enable further enhancements, creating a cycle of continuous self-enhancement.
140
00:15:13,795 --> 00:15:23,285
Historically, the aim has been to make progress cumulative, ensuring that each advancement not only enhances the system but also its ability to self-improve.
141
00:15:23,333 --> 00:15:27,973
Evan: Can you give us examples of early approaches to RSI?
142
00:15:28,037 --> 00:15:28,927
Ashley: Sure.
143
00:15:28,927 --> 00:15:34,707
Early methods included self-feedback mechanisms and the retention of reflections for subsequent attempts.
144
00:15:34,707 --> 00:15:41,937
For instance, models like self-rewarding language models use their own outputs as feedback to adjust their strategies over time.
145
00:15:41,937 --> 00:15:49,837
Additionally, frameworks like STOP (Self-Taught Optimizer Program) focus on recursively rewriting code-improvement programs.
146
00:15:49,941 --> 00:15:54,501
Evan: And how does Harness2 differentiate itself from these earlier methods?
147
00:15:54,557 --> 00:16:03,097
Ashley: The key difference with Harness2 lies in its emphasis on the harnesses surrounding a frozen model rather than adjusting the model weights themselves.
148
00:16:03,097 --> 00:16:13,167
Essentially, Harness2 focuses on optimizing the execution context, including prompts, tools, and orchestration, as opposed to just tweaking the model parameters.
149
00:16:13,167 --> 00:16:20,577
It uses a contrastive improvement process to iteratively refine the harness components and accumulate reusable experience.
150
00:16:20,621 --> 00:16:26,981
Evan: That leads me to wonder, what are
some of the recent developments in harness optimization and evolution?
151
00:16:27,029 --> 00:16:32,969
Ashley: Recent efforts in this area have expanded beyond prompts and workflows to include comprehensive harnesses.
152
00:16:32,969 --> 00:16:39,209
Some notable methods include searching over candidate harnesses using training tasks and evaluation feedback.
153
00:16:39,209 --> 00:16:50,049
For example, Meta-Harness searches over harnesses offline to find optimal configurations, while other methods use task rewards or diagnostic experience to guide the harness changes.
154
00:16:50,093 --> 00:16:52,663
Evan: It's fascinating to see this evolution.
155
00:16:52,663 --> 00:16:54,113
What about online methods?
156
00:16:54,113 --> 00:16:57,273
Are there approaches that refine harnesses in real-time?
157
00:16:57,317 --> 00:17:02,827
Ashley: Yes, online adaptation methods refine harnesses during actual task executions.
158
00:17:02,827 --> 00:17:07,287
These methods allow the system to adapt dynamically to new tasks as they arrive.
159
00:17:07,287 --> 00:17:15,867
For instance, task-specific harnesses can evolve by incorporating feedback from the current task to improve their configurations for subsequent tasks.
160
00:17:15,867 --> 00:17:21,877
This continuous adaptation helps the system manage the unpredictable nature of open-world settings.
161
00:17:22,011 --> 00:17:28,601
Evan: Could you elaborate on some of the systems that combine online adaptation with ongoing evaluation?
162
00:17:28,661 --> 00:17:29,581
Ashley: Certainly.
163
00:17:29,581 --> 00:17:37,341
For example, systems like MemoHarness build a diagnostic experience bank that supports case-specific edits during execution.
164
00:17:37,341 --> 00:17:44,781
Others, like Evo-Harness, focus on retaining improvement state across task cycles, thus learning from delayed outcome feedback.
165
00:17:44,781 --> 00:17:56,941
Harness2, however, stands out by combining contrastive task-level refinement with harnessing-level recursion, enabling it to learn from a continual stream of tasks without needing explicit external supervision.
166
00:17:57,065 --> 00:18:02,615
Evan: It seems like retaining and leveraging past experiences is a crucial component here.
167
00:18:02,615 --> 00:18:05,925
How does this compare with the traditional static approaches?
168
00:18:05,981 --> 00:18:13,331
Ashley: Static approaches, which select and deploy a fixed harness, often struggle to generalize to new, unforeseen tasks.
169
00:18:13,331 --> 00:18:21,361
In contrast, recursive approaches like Harness2 evaluate and refine harnesses continually, which makes them more adaptable.
170
00:18:21,361 --> 00:18:28,841
Each new task contributes to a growing repository of experience, enhancing the system’s ability to deal with future challenges.
171
00:18:28,901 --> 00:18:35,661
Evan: And where does Harness2 stand in terms of its novelty and impact compared to existing methods?
172
00:18:35,747 --> 00:18:43,847
Ashley:
Harness2 brings a novel perspective by separating the execution harness from the improvement harness and systematically refining both levels.
173
00:18:43,847 --> 00:18:52,197
This dual-recursion mechanism allows the system to not only improve its current task performance but also become better at future harnessing efforts.
174
00:18:52,197 --> 00:19:00,477
The cumulative lessons stored in the improvement library provide a valuable asset, setting the stage for more efficient and accurate harness refinements.
175
00:19:00,593 --> 00:19:08,553
Evan: So, Harness2 is essentially creating a foundation for more sophisticated and adaptable AI systems moving forward?
176
00:19:08,597 --> 00:19:09,677
Ashley: Exactly.
177
00:19:09,677 --> 00:19:17,797
By focusing on recursive improvements at both the task and harnessing levels, Harness2 aims to build a more robust and flexible system.
178
00:19:17,797 --> 00:19:26,137
Its methodologies and findings contribute significantly to the ongoing discourse on optimizing AI systems for open-world applications.
179
00:19:26,189 --> 00:19:28,599
Evan: That wraps up the related work section.
180
00:19:28,599 --> 00:19:31,609
Next, we will dive into the paper's conclusions.
181
00:19:32,862 --> 00:19:43,402
Evan: Let's move on to the final part of our podcast, where we summarize the key contributions and takeaways from the paper 'Harness2: Recursive Agent Harnessing for an Open World'.
182
00:19:43,476 --> 00:19:44,376
Ashley: Evan.
183
00:19:44,376 --> 00:19:52,416
This paper introduces Harness2, a framework addressing the significant challenges of recursive agent harnessing in open-world environments.
184
00:19:52,416 --> 00:20:03,746
By employing two levels of recursion—task-level and harnessing-level—Harness2 refines the agent's execution harness and learns from each iteration to improve future harnessing efforts.
185
00:20:03,798 --> 00:20:11,418
Evan: And at the core of this framework is the innovative Propose–Probe–Compose step, right?
186
00:20:11,478 --> 00:20:12,458
Ashley: Exactly.
187
00:20:12,458 --> 00:20:25,828
The Propose–Probe–Compose step forms the backbone of task-level recursion, enabling the system to propose, test, and refine candidate harness edits based on behavioral contrasts rather than ground-truth verifiers.
188
00:20:25,828 --> 00:20:31,278
This process ensures reliable harness improvements even in the absence of definitive labels.
189
00:20:31,326 --> 00:20:35,676
Evan: Harness2 also continuously updates a harness improvement library.
190
00:20:35,676 --> 00:20:37,906
How does this benefit future tasks?
191
00:20:37,950 --> 00:20:45,650
Ashley: The harness improvement library stores detailed records of effective harness changes along with contextual evidence of their performance.
192
00:20:45,650 --> 00:20:57,150
This repository of reusable components and lessons learned enables the system to efficiently adapt to new tasks, continually building on its prior experience without redundant probing.
193
00:20:57,198 --> 00:21:00,538
Evan: What were the key findings from their experiments?
194
00:21:00,582 --> 00:21:16,252
Ashley: Harness2 demonstrated consistent improvements across multiple benchmarks and configurations, with gains up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB's Partial pass rate.
195
00:21:16,252 --> 00:21:22,142
The results highlight the scalability and effectiveness of both parallel and sequential recursion strategies.
196
00:21:22,246 --> 00:21:33,746
Evan: It’s evident that Harness2 offers a robust solution for evolving agent harnesses dynamically while leveraging past experiences for continuous improvement.
197
00:21:33,798 --> 00:21:35,068
Ashley: Indeed, Evan.
198
00:21:35,068 --> 00:21:49,298
By addressing both current task requirements and preparing for future adaptability, Harness2 sets a new standard in recursive self-improvement for AI systems, making significant strides in tackling the unpredictability of open-world settings.
199
00:21:49,350 --> 00:21:51,160
Evan: That’s our episode for today.
200
00:21:51,160 --> 00:21:55,130
We hope you found this discussion on Harness2 insightful.
201
00:21:55,182 --> 00:22:00,952
Ashley: Join us next time as we continue to break down the latest advancements in AI research.
202
00:22:00,952 --> 00:22:05,862
Keep tuning in to Daily Paper Cast for your daily dose of AI knowledge.
203
00:22:05,910 --> 00:22:09,870
Evan: Thanks for listening, and see you in the next episode!
204
00:22:09,870 --> 00:22:11,050
Bye for now.
205
00:22:11,124 --> 00:22:12,714
Ashley: Goodbye everyone!