MAROKO133 Update ai: Which Agent Causes Task Failures and When?Researchers from PSU and Du

📌 MAROKO133 Eksklusif ai: Which Agent Causes Task Failures and When?Researchers fr

Share My Research is Synced’s column that welcomes scholars to share their own research breakthroughs with over 1.5M global AI enthusiasts. Beyond technological advances, Share My Research also calls for interesting stories behind the research and exciting research ideas. Contact us: [email protected]

Meet the authors
Institutions: Penn State University, Duke University, Google DeepMind, University of Washington, Meta, Nanyang Technological University, and Oregon State University. The co-first authors are Shaokun Zhang of Penn State University and Ming Yin of Duke University.

In recent years, LLM Multi-Agent systems have garnered widespread attention for their collaborative approach to solving complex problems. However, it’s a common scenario for these systems to fail at a task despite a flurry of activity. This leaves developers with a critical question: which agent, at what point, was responsible for the failure? Sifting through vast interaction logs to pinpoint the root cause feels like finding a needle in a haystack—a time-consuming and labor-intensive effort.
 
This is a familiar frustration for developers. In increasingly complex Multi-Agent systems, failures are not only common but also incredibly difficult to diagnose due to the autonomous nature of agent collaboration and long information chains. Without a way to quickly identify the source of a failure, system iteration and optimization grind to a halt.
 
To address this challenge, researchers from Penn State University and Duke University, in collaboration with institutions including Google DeepMind, have introduced the novel research problem of “Automated Failure Attribution.” They have constructed the first benchmark dataset for this task, Who&When, and have developed and evaluated several automated attribution methods. This work not only highlights the complexity of the task but also paves a new path toward enhancing the reliability of LLM Multi-Agent systems.

The paper has been accepted as a Spotlight presentation at the top-tier machine learning conference, ICML 2025, and the code and dataset are now fully open-source.

Paper：https://arxiv.org/pdf/2505.00212
Code：https://github.com/mingyin1/Agents_Failure_Attribution
Dataset：https://huggingface.co/datasets/Kevin355/Who_and_When
 
 
Research Background and Challenges
LLM-driven Multi-Agent systems have demonstrated immense potential across many domains. However, these systems are fragile; errors by a single agent, misunderstandings between agents, or mistakes in information transmission can lead to the failure of the entire task.

Currently, when a system fails, developers are often left with manual and inefficient methods for debugging:
Manual Log Archaeology : Developers must manually review lengthy interaction logs to find the source of the problem.
Reliance on Expertise : The debugging process is highly dependent on the developer’s deep understanding of the system and the task at hand.
 
This “needle in a haystack” approach to debugging is not only inefficient but also severely hinders rapid system iteration and the improvement of system reliability. There is an urgent need for an automated, systematic method to pinpoint the cause of failures, effectively bridging the gap between “evaluation results” and “system improvement.”

Core Contributions
This paper makes several groundbreaking contributions to address the challenges above:
1. Defining a New Problem: The paper is the first to formalize “automated failure attribution” as a specific research task. This task is defined by identifying the failure-responsible agent and the decisive error step that led to the task’s failure.
2. Constructing the First Benchmark Dataset: Who&When : This dataset includes a wide range of failure logs collected from 127 LLM Multi-Agent systems, which were either algorithmically generated or hand-crafted by experts to ensure realism and diversity. Each failure log is accompanied by fine-grained human annotations for:
Who: The agent responsible for the failure.
When: The specific interaction step where the decisive error occurred.
Why: A natural language explanation of the cause of the failure.

3. Exploring Initial “Automated Attribution” Methods : Using the Who&When dataset, the paper designs and assesses three distinct methods for automated failure attribution:
– All-at-Once: This method provides the LLM with the user query and the complete failure log, asking it to identify the responsible agent and the decisive error step in a single pass. While cost-effective, it may struggle to pinpoint precise errors in long contexts.
– Step-by-Step: This approach mimics manual debugging by having the LLM review the interaction log sequentially, making a judgment at each step until the error is found. It is more precise at locating the error step but incurs higher costs and risks accumulating errors.
– Binary Search: A compromise between the first two methods, this strategy repeatedly divides the log in half, using the LLM to determine which segment contains the error. It then recursively searches the identified segment, offering a balance of cost and performance.

Experimental Results and Key Findings 
Experiments were conducted in two settings: one where the LLM knows the ground truth answer to the problem the Multi-Agent system is trying to solve (With Ground Truth) and one where it does not (Without Ground Truth). The primary model used was GPT-4o, though other models were also tested. The systematic evaluation of these methods on the Who&When dataset yielded several important insights:
– A Long Way to Go: Current methods are far from perfect. Even the best-performing single method achieved an accuracy of only about 53.5% in identifying the responsible agent and a mere 14.2% in pinpointing the exact error step. Some methods performed even worse than random guessing, underscoring the difficulty of the task.
– No “All-in-One” Solution: Different methods excel at different aspects of the problem. The All-at-Once method is better at identifying “Who,” while the Step-by-Step method is more effective at determining “When.” The Binary Search method provides a middle-ground performance.
 

– Hybrid Approaches Show Promise but at a High Cost: The researchers found that combining different methods, such as using the All-at-Once approach to identify a potential agent and then applying the Step-by-Step method to find the error, can improve overall performance. However, this comes with a significant increase in computational cost.

– State-of-the-Art Models Struggle: Surprisingly, even the most advanced reasoning models, like OpenAI o1 and DeepSeek R1, find this task challenging.- This h…

Konten dipersingkat otomatis.

🔗 Sumber: syncedreview.com

📌 MAROKO133 Hot ai: Nous Research just released Nomos 1, an open-source AI that ra

Nous Research, the San Francisco-based artificial intelligence startup, released on Tuesday an open-source mathematical reasoning system called Nomos 1 that achieved near-elite human performance on this year's William Lowell Putnam Mathematical Competition, one of the most prestigious and notoriously difficult undergraduate math contests in the world.

The Putnam is known for its difficulty: While a perfect score is 120, this year's top score was 90, and the median was just 2. Nomos 1, by contrast, scored 87 points — a result that would have ranked second out of 3,988 participants in the 2024 competition, according to the company.

The release marks an inflection point in the rapidly accelerating race to build AI systems capable of sophisticated mathematical reasoning. Unlike the massive, compute-intensive models deployed by major technology companies, Nomos 1 achieves its results with a relatively compact architecture: 30 billion parameters with roughly 3 billion active at any given time, using a mixture-of-experts design based on Alibaba's Qwen3 model.

"This score would rank #2/3988 in 2024 and marks our first step with Hillclimb AI towards creating a SOTA AI mathematician," Nous Research announced on social media Tuesday.

The same base model scored 24 points without Nous Research's specialized training

Perhaps most striking is the gap between Nomos 1 and its base model. When Nous Research ran the same Qwen3-30B-A3B-Thinking-2507 model through an identical testing harness, it scored just 24 out of 120 — a result that underscores the critical importance of post-training optimization and specialized reasoning techniques over raw model scale.

"Nomos 1 achieved an 87/120 with 8 perfect scores," the company stated, noting that the performance difference "is largely due to post-training and data quality rather than the harness."

The results were verified through blind grading by a human expert who had previously finished in the top 200 on the Putnam. Nous Research provided the anonymized submissions to the grader, then published the full set of de-anonymized files and the runbooks used to generate them on GitHub.

Why the Putnam competition is considered the ultimate test of mathematical reasoning

The William Lowell Putnam Mathematical Competition is an annual mathematics competition for undergraduate college students enrolled at institutions of higher learning in the United States and Canada. It is widely considered to be the most prestigious university-level mathematics competition in the world.

The notoriously brutal William Lowell Putnam Mathematical Competition is more of a mathematical sporting event than an academic test. The exam consists of two 3-hour sessions separated by a 2-hour break. There are a total of 12 questions to be solved, 6 for each session. Each question is worth 10 points, for a total of 120 points.

Putnam questions are not the type that come up in regular exams or textbooks. They are more like puzzles than calculations, often requiring students to find different ways to represent things before a solution might unfold.

Last year, nearly 4,000 students across the continent wrote the Putnam. Sixty-one per cent scored three points or fewer, according to the Mathematical Association of America, which organizes the competition. The top score was 90 out of 120.

Many Putnam Fellows have gone on to become distinguished researchers in mathematics and other fields, including three Fields Medalists — John Milnor, David Mumford, and Daniel Quillen — and two Nobel laureates in physics — Richard Feynman and Kenneth Wilson.

Inside the two-phase reasoning system that powers Nomos 1's mathematical breakthroughs

Nomos 1 is a specialization of Qwen's Qwen3-30B-A3B-Thinking model, optimized for mathematical problem-solving and proof-writing in natural language. The system was developed in collaboration with Hillclimb AI.

What distinguishes Nomos 1 from simple model inference is its sophisticated reasoning harness — an open-source framework that orchestrates how the model approaches and solves problems. The harness operates in two distinct phases within a three-hour time limit, mirroring the actual Putnam competition structure.

In the solving phase, parallel workers simultaneously tackle problems using a priority-based system. Each worker picks a problem, generates a submission, then scores its own work on a scale of 1 to 7. Problems with the fewest perfect scores receive priority, ensuring the system focuses its compute on the hardest challenges. This process continues until either all problems have achieved a target number of self-critiqued perfect scores or time runs out.

The finalization phase begins 15 minutes before the time limit (or at 50% for shorter runs) and employs a two-stage selection process. First, a consolidation step groups submissions by conclusion and attempts to identify the correct group — importantly, not necessarily the majority group. Then, a pairwise tournament using single elimination determines the final submission for each problem.

"Our open source reasoning system consists of a solving phase, where workers attempt a least-solved problem and self-assess, followed by a finalization phase, which consolidates submissions to choose a final submission for each problem," Nous Research explained.

How Nomos 1 compares to mathematical AI systems from DeepSeek, Google, and OpenAI

The Nomos 1 results arrive amid a flurry of advances in mathematical reasoning AI. DeepSeek's model, DeepSeekMath-V2, scored 118 out of 120 points on questions from the 2024 William Lowell Putnam Mathematical Competition, beating the top human score of 90. The model also performed at the level of gold-medal winners in the International Mathematical Olympiad.

This year, Google's advanced Gemini model operated end-to-end in natural language, producing rigorous mathematical proofs directly from the official problem descriptions – all within the 4.5-hour competition time limit. They achieved this year's result using an advanced version of Gemini Deep Think.

What makes Nomos 1's achievement notable is not raw performance — it trails DeepSeek's 118/120 — but rather its accessibility and efficiency. At 30 billion parameters with only 3 billion active, the model can run on consumer-grade hardware, a stark contrast to the massive compute clusters required by frontier models from OpenAI and Google.

Hermes 4.3 arrived just six days earlier, trained on a decentralized blockchain network

The Nomos 1 anno…

Konten dipersingkat otomatis.

🔗 Sumber: venturebeat.com