Can AI Truly Do Science?

Author: Denis Avetisyan

A new approach uses artificially generated research problems to train artificial intelligence agents in the iterative process of scientific discovery.

The system autonomously generates machine learning tasks and datasets without human intervention, employing a debugging loop to resolve compilation issues and refine the generated tasks rather than abandoning them outright.

Researchers present a scalable pipeline for synthetic task generation, enabling AI to perform multi-hop reasoning and long-form question answering for automated machine learning research.

Despite advances in artificial intelligence, training agents to autonomously conduct meaningful scientific research remains a significant challenge, often hindered by a lack of principled training data. This work, ‘AI Scientist via Synthetic Task Scaling’, introduces a novel pipeline for automatically generating scalable and high-quality synthetic machine learning research tasks, complete with dataset proposals and code generation. Experiments demonstrate that training AI agents on these synthetic tasks-verified against real-world data via the Huggingface API-significantly improves performance on benchmark MLGym tasks, raising the AUP metric by up to 12% for tested models. Could this approach pave the way for AI systems capable of independent scientific discovery and iterative problem-solving in complex domains?

The Challenge of Complex Information Synthesis

Contemporary question answering systems face significant hurdles when tasked with long-form question answering (LongFormQA), a process demanding the synthesis of information dispersed across multiple source documents. Unlike systems designed to retrieve answers from a single text passage, LongFormQA requires a deeper level of comprehension and the ability to establish connections between disparate pieces of information. This presents a considerable challenge, as current models often struggle to effectively aggregate and reconcile conflicting or nuanced details presented in multiple documents, leading to incomplete or inaccurate responses. The difficulty isn’t simply locating relevant information, but rather in performing the complex reasoning necessary to construct a cohesive and well-supported answer from a fragmented knowledge base, a key limitation in achieving truly intelligent question answering capabilities.

Current question answering systems frequently encounter difficulties when tasked with synthesizing information from multiple sources, a critical limitation impacting their ability to address complex queries. The challenge lies not simply in retrieving relevant passages, but in effectively integrating these disparate pieces of information into a coherent and accurate response. These systems often struggle to identify relationships between facts presented in different documents, leading to fragmented or contradictory answers. This inability to perform robust information synthesis hinders performance on long-form question answering, where a complete understanding necessitates a holistic view derived from multiple documents, rather than isolated facts. Consequently, even with access to vast knowledge bases, the systems’ responses can lack the nuance and depth required for truly insightful answers.

HotpotQA: A Rigorous Testbed for Multi-Hop Reasoning

The HotpotQA dataset is designed to assess an AI agent’s capacity for multi-hop reasoning, requiring the synthesis of information from multiple supporting documents to answer a question. Unlike simpler question answering datasets, HotpotQA necessitates identifying several relevant passages – typically more than one – and then logically combining the information contained within them. The dataset consists of questions posed over long documents, with answers requiring inference rather than direct retrieval. Evaluation metrics focus on both answer accuracy and the ability to correctly identify the supporting facts used to arrive at the answer, providing a comprehensive assessment of reasoning capabilities beyond simple fact matching.

The HotpotQA dataset incorporates a DistractorSetting to specifically challenge question answering models with extraneous information. This setting includes irrelevant documents and sentences alongside those containing supporting facts, requiring models to discern which information is genuinely pertinent to answering the given question. The inclusion of distractors increases the difficulty of the task by forcing models to move beyond simple keyword matching and instead perform a more nuanced evaluation of semantic relevance. Performance within the DistractorSetting serves as a key indicator of a model’s ability to filter noise and focus on the critical evidence needed for accurate multi-hop reasoning.

Robust Supporting Fact Selection is essential for performance in question answering tasks involving distractor settings, such as the HotpotQA dataset. This process requires identifying and extracting only the sentences directly relevant to answering the question, while effectively filtering out irrelevant or contradictory information presented as distractors. Models must not only locate supporting facts within a document but also discern their validity and relationship to the query, demanding a nuanced understanding of semantic relevance beyond simple keyword matching. Failure to accurately select supporting facts leads to incorrect answers, even when the necessary information is present within the provided context.

Establishing a Baseline for Answer Extraction

A BaselineModel in answer extraction functions as a preliminary system against which the performance of more complex models is evaluated. This foundational approach typically employs relatively simple algorithms – such as keyword matching or regular expressions – to identify potential answers within a given text. Establishing this baseline is critical for determining whether subsequent, more computationally intensive techniques offer statistically significant improvements in accuracy, recall, or F1-score. The BaselineModel’s primary purpose isn’t to achieve state-of-the-art results, but rather to provide a consistent and easily reproducible standard for comparative analysis during the development and refinement of answer extraction systems.

AnswerExtraction is the core component enabling a BaselineModel to function; it involves locating specific segments, or spans, of text within a source document that directly address the posed question. This process typically involves analyzing the question’s semantic content to identify relevant keywords and entities, then searching the document for matching or related textual units. The identified spans are then evaluated based on factors such as contextual relevance and grammatical correctness to determine the most appropriate answer. The output of AnswerExtraction is a discrete textual segment representing the model’s response to the input question, forming the basis for performance evaluation and comparison with other extraction methods.

The BaselineModel frequently utilizes JSONFormat for data input due to its inherent structure and parsing efficiency. JSON’s key-value pairs allow for clear delineation of question-answer relationships and document context, facilitating automated processing. This format enables the model to readily access relevant information, including the question text, the document containing the answer, and the character-based offsets indicating the answer’s location within the document. The use of JSON also simplifies data validation and integration with various programming languages and machine learning frameworks, streamlining the development and deployment of the BaselineModel.

Our supervised fine-tuned (SFT) Qwen3 models outperform the base Qwen3-4B model on the majority of MLGym subtasks, as demonstrated by aggregated violin plots across 64 runs, indicating improved performance in 9 out of 13 tasks.

Measuring Performance: Comprehensive Evaluation Metrics

Question answering systems are rigorously evaluated using metrics designed to quantify both accuracy and completeness. Exact Match (EM) determines if a model’s generated answer perfectly matches a known correct answer, offering a strict measure of precision. However, recognizing that multiple valid answers often exist, the F1 score provides a more nuanced assessment by calculating the harmonic mean of precision and recall between the generated and reference answers. This metric considers overlapping words and phrases, acknowledging partial correctness even when an exact match isn’t achieved. By employing both EM and F1, researchers gain a comprehensive understanding of a model’s ability to not only find the right answer, but also to provide complete and relevant information, crucial for building trustworthy and effective question answering systems.

A thorough evaluation of question answering systems necessitates assessing performance across multiple facets of the process, and consequently, metrics like Exact Match and F1 score aren’t solely applied to the final answer. These measures are also critically used to evaluate SupportingFactSelection, the model’s ability to correctly identify the relevant evidence from a knowledge source that justifies its answer. By independently scoring both AnswerExtraction – the precision of the answer itself – and SupportingFactSelection, researchers gain a nuanced understanding of where a model excels or falters. This dual assessment reveals whether a poor answer stems from an inability to find the right information, or from a failure to synthesize that information correctly, ultimately enabling more targeted improvements to the system’s architecture and training data.

Evaluations on the MLGym benchmark reveal a substantial enhancement in performance facilitated by this approach. Specifically, question answering models utilizing the developed method achieved a noteworthy 9% gain when tested on the Qwen3-4B architecture, and an even more pronounced 12% improvement with the larger Qwen3-8B model. These results indicate a clear and significant advancement over existing baseline models, demonstrating the effectiveness of the techniques employed in enhancing the accuracy and reliability of multi-hop reasoning and answer extraction capabilities.

A crucial element in assessing the capacity of question answering systems to perform complex reasoning is the Joint F1 score, a metric specifically designed to evaluate performance on multi-hop questions-those requiring synthesis of information from multiple sources. Measurements revealed a Joint F1 score of 0.022210986997935424, indicating the effectiveness of the presented method in accurately connecting disparate pieces of information to arrive at correct answers. This score signifies a substantial capability in navigating the complexities of multi-hop reasoning tasks, where simply retrieving relevant facts is insufficient; the system must also logically integrate them. The resulting score provides quantitative evidence supporting the method’s proficiency in tackling questions that demand more than superficial information retrieval.

The presented research embodies a philosophy of incremental development, mirroring the evolution of complex systems. Just as infrastructure should evolve without rebuilding the entire block, this pipeline for synthetic task scaling allows for iterative refinement of AI agents’ scientific capabilities. The generation of increasingly complex multi-hop reasoning challenges, facilitated by synthetic data, demonstrates a commitment to building upon existing foundations rather than attempting wholesale redesign. This approach aligns with Dijkstra’s observation that, “It is not enough to make things working; they must also be understandable.” The pipeline’s scalability and focus on verifiable experimentation promote not only functional progress but also a deeper comprehension of the AI’s learning process – a crucial element for robust and reliable scientific discovery.

Beyond the Horizon

The presented work establishes a method for generating complexity, but does not resolve the fundamental question of what constitutes meaningful scientific progress. Scaling synthetic tasks reveals the capacity for automated discovery, yet the true measure lies in the novelty and generalizability of those discoveries – qualities difficult to assess within a self-contained system. The elegance of this pipeline rests on its ability to sidestep the bottlenecks of human-curated datasets, but it simultaneously inherits the biases embedded within the initial generative models. A crucial next step involves mechanisms for evaluating the ‘surprisingness’ of results, and actively steering the agent towards unexplored regions of the scientific landscape.

One anticipates that future iterations will focus on refining the feedback loops – not simply rewarding successful experimentation, but penalizing unproductive paths and encouraging conceptual leaps. The current framework treats the agent as a solitary explorer; integrating multiple agents, each with specialized roles and competing hypotheses, could foster a more robust and efficient discovery process. However, such a system introduces new challenges regarding coordination, communication, and the potential for emergent, and perhaps undesirable, behaviors.

Ultimately, the limitations are not computational, but conceptual. The real frontier lies in defining what it means to ‘understand’ a scientific phenomenon, and translating that understanding into actionable insights. A scalable pipeline is merely a tool; the quality of the science it produces depends entirely on the clarity and depth of the questions it is designed to address.

Original article: https://arxiv.org/pdf/2603.17216.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

The Challenge of Complex Information Synthesis

HotpotQA: A Rigorous Testbed for Multi-Hop Reasoning

Establishing a Baseline for Answer Extraction

Measuring Performance: Comprehensive Evaluation Metrics

Beyond the Horizon

See also: