Lawrence Phillips

Peter Mühlbacher, Nikos I. Bosse, Lawrence Phillips

We present initial results of a forthcoming benchmark for evaluating LLM agents on white-collar tasks of economic value. We evaluate agents on real-world "messy" open-web research tasks of the type that are routine in finance and consulting. In doing so, we lay the groundwork for an LLM agent evaluation suite where good performance directly corresponds to a large economic and societal impact. We built and tested several agent architectures with o1-preview, GPT-4o, Claude-3.5 Sonnet, Llama 3.1 (405b), and GPT-4o-mini. On average, LLM agents powered by Claude-3.5 Sonnet and o1-preview substantially outperformed agents using GPT-4o, with agents based on Llama 3.1 (405b) and GPT-4o-mini lagging noticeably behind. Across LLMs, a ReAct architecture with the ability to delegate subtasks to subagents performed best. In addition to quantitative evaluations, we qualitatively assessed the performance of the LLM agents by inspecting their traces and reflecting on their observations. Our evaluation represents the first in-depth assessment of agents' abilities to conduct challenging, economically valuable analyst-style research on the real open web.

Nikos I. Bosse, Peter Mühlbacher, Jack Wildman et al.

Forecasting future events is highly valuable in decision-making and is a robust measure of general intelligence. As forecasting is probabilistic, developing and evaluating AI forecasters requires generating large numbers of diverse and difficult questions, and accurately resolving them. Previous efforts to automate this laborious work relied on recurring data sources (e.g., weather, stocks), limiting diversity and utility. In this work, we present a system for generating and resolving high-quality forecasting questions automatically and at scale using LLM-powered web research agents. We use this system to generate 1499 diverse, real-world forecasting questions, and to resolve them several months later. We estimate that our system produces verifiable, unambiguous questions approximately 96% of the time, exceeding the rate of Metaculus, a leading human-curated forecasting platform. We also find that our system resolves questions at approximately 95% accuracy. We verify that forecasting agents powered by more intelligent LLMs perform better on these questions (Brier score of 0.134 for Gemini 3 Pro, 0.149 for GPT-5, and 0.179 for Gemini 2.5 Flash). Finally, we demonstrate how our system can be leveraged to directly improve forecasting, by evaluating a question decomposition strategy on a generated question set, yielding a significant improvement in Brier scores (0.132 vs. 0.141).

FutureSearch, Nikos I. Bosse, Jon Evans et al.

Amongst the most common use cases of modern AI is LLM chat with web search enabled. However, no direct evaluations of the quality of web research agents exist that control for the continually-changing web. We introduce Deep Research Bench, consisting of 89 multi-step web research task instances of varying difficulty across 8 diverse task categories, with the answers carefully worked out by skilled humans. We provide a "RetroSearch" environment with a large frozen set of scraped web pages, and demonstrate that offline "RetroSearch" agents perform comparably to "live web" agents, enabling reliable evaluations of models over time. We provide robust agent tooling and scaffolding to benchmark major LLMs as they are released, including "thinking" models like o3 and Gemini 2.5 Pro. We include automated evaluations of the lengthy agent traces to report progress over time in hallucinations, tool use, and forgetting. Finally, we evaluate the major web research products branded as "Deep Research", "Deep Search", "Search", or "Research." Results are available on a public leaderboard at https://drb.futuresearch.ai/.

FutureSearch, Jack Wildman, Nikos I. Bosse et al.

Forecasting is a challenging task that offers a clearly measurable way to study AI systems. Forecasting requires a large amount of research on the internet, and evaluations require time for events to happen, making the development of forecasting benchmarks challenging. To date, no forecasting benchmark provides a realistic, hermetic, and repeatable environment for LLM forecasters. We introduce Bench To the Future (BTF), a "pastcasting" benchmark with hundreds of high-quality questions for which the resolution is already known. Each question is accompanied by a large offline corpus of tens of thousands of relevant web pages, enabling a way to elicit realistic "forecasts" on past events from LLMs. Results suggest that our pastcasting environment can produce results comparable to those based on forecasts using the internet on at-the-time unresolved questions. We show results benchmarking agent and chain-of-thought forecasting approaches using several LLMs, including the recently-released Claude 4 models, and demonstrate BTF's ability to track steady forecasting capability progress over time. We intend this to be a living benchmark, with new questions added continually to account for increasing training data cutoff dates. We invite researchers to contact us at hello@futuresearch.ai to utilize our benchmark or tooling for their own research.

Josefa Lia Stoisser, Marc Boubnovski Martell, Lawrence Phillips et al.

Large language model (LLM) agents are increasingly deployed in structured biomedical data environments, yet they often produce fluent but overconfident outputs when reasoning over complex multi-table data. We introduce an uncertainty-aware agent for query-conditioned multi-table summarization that leverages two complementary signals: (i) retrieval uncertainty--entropy over multiple table-selection rollouts--and (ii) summary uncertainty--combining self-consistency and perplexity. Summary uncertainty is incorporated into reinforcement learning (RL) with Group Relative Policy Optimization (GRPO), while both retrieval and summary uncertainty guide inference-time filtering and support the construction of higher-quality synthetic datasets. On multi-omics benchmarks, our approach improves factuality and calibration, nearly tripling correct and useful claims per summary (3.0\(\rightarrow\)8.4 internal; 3.6\(\rightarrow\)9.9 cancer multi-omics) and substantially improving downstream survival prediction (C-index 0.32\(\rightarrow\)0.63). These results demonstrate that uncertainty can serve as a control signal--enabling agents to abstain, communicate confidence, and become more reliable tools for complex structured-data environments.

Lawrence Phillips, Marc Boubnovski Martell, Aditya Misra et al.

Predicting cellular responses to genetic perturbations represents a fundamental challenge in systems biology, critical for advancing therapeutic discovery and virtual cell modeling. While large language models (LLMs) show promise for biological reasoning, their application to perturbation prediction remains underexplored due to challenges in adapting them to structured experimental data. We present SynthPert, a novel method that enhances LLM performance through supervised fine-tuning on synthetic reasoning traces generated by frontier models. Using the PerturbQA benchmark, we demonstrate that our approach not only achieves state-of-the-art performance but surpasses the capabilities of the frontier model that generated the training data. Our results reveal three key insights: (1) Synthetic reasoning traces effectively distill biological knowledge even when partially inaccurate, (2) This approach enables cross-cell-type generalization with 87% accuracy on unseen RPE1 cells, and (3) Performance gains persist despite using only 2% of quality-filtered training data. This work shows the effectiveness of synthetic reasoning distillation for enhancing domain-specific reasoning in LLMs.

Francesco Farina, Lawrence Phillips, Nicola J Richmond

We introduce a framework for uncertainty estimation that both describes and extends many existing methods. We consider typical hyperparameters involved in classical training as random variables and marginalise them out to capture various sources of uncertainty in the parameter space. We investigate which forms and combinations of marginalisation are most useful from a practical point of view on standard benchmarking data sets. Moreover, we discuss how some marginalisations may produce reliable estimates of uncertainty without the need for extensive hyperparameter tuning and/or large-scale ensembling.

Lawrence Phillips, Garrett Goh, Nathan Hodas

Neural network interpretability is a vital component for applications across a wide variety of domains. In such cases it is often useful to analyze a network which has already been trained for its specific purpose. In this work, we develop a method to produce explanation masks for pre-trained networks. The mask localizes the most important aspects of each input for prediction of the original network. Masks are created by a secondary network whose goal is to create as small an explanation as possible while still preserving the predictive accuracy of the original network. We demonstrate the applicability of our method for image classification with CNNs, sentiment analysis with RNNs, and chemical property prediction with mixed CNN/RNN architectures.

Chris Careaga, Brian Hutchinson, Nathan Hodas et al.

In the few-shot scenario, a learner must effectively generalize to unseen classes given a small support set of labeled examples. While a relatively large amount of research has gone into few-shot learning for image classification, little work has been done on few-shot video classification. In this work, we address the task of few-shot video action recognition with a set of two-stream models. We evaluate the performance of a set of convolutional and recurrent neural network video encoder architectures used in conjunction with three popular metric-based few-shot algorithms. We train and evaluate using a few-shot split of the Kinetics 600 dataset. Our experiments confirm the importance of the two-stream setup, and find prototypical networks and pooled long short-term memory network embeddings to give the best performance as few-shot method and video encoder, respectively. For a 5-shot 5-way task, this setup obtains 84.2% accuracy on the test set and 59.4% on a special "challenge" test set, composed of highly confusable classes.

Matthias Bal, Hagen Triendl, Mariana Assmann et al.

Architectures for sparse hierarchical representation learning have recently been proposed for graph-structured data, but so far assume the absence of edge features in the graph. We close this gap and propose a method to pool graphs with edge features, inspired by the hierarchical nature of chemistry. In particular, we introduce two types of pooling layers compatible with an edge-feature graph-convolutional architecture and investigate their performance for molecules relevant to drug discovery on a set of two classification and two regression benchmark datasets of MoleculeNet. We find that our models significantly outperform previous benchmarks on three of the datasets and reach state-of-the-art results on the fourth benchmark, with pooling improving performance for three out of four tasks, keeping performance stable on the fourth task, and generally speeding up the training process.

Nathan Hilliard, Lawrence Phillips, Scott Howland et al.

Learning high quality class representations from few examples is a key problem in metric-learning approaches to few-shot learning. To accomplish this, we introduce a novel architecture where class representations are conditioned for each few-shot trial based on a target image. We also deviate from traditional metric-learning approaches by training a network to perform comparisons between classes rather than relying on a static metric comparison. This allows the network to decide what aspects of each class are important for the comparison at hand. We find that this flexible architecture works well in practice, achieving state-of-the-art performance on the Caltech-UCSD birds fine-grained classification task.

Lawrence Phillips, Nathan Hodas

Increasingly, cognitive scientists have demonstrated interest in applying tools from deep learning. One use for deep learning is in language acquisition where it is useful to know if a linguistic phenomenon can be learned through domain-general means. To assess whether unsupervised deep learning is appropriate, we first pose a smaller question: Can unsupervised neural networks apply linguistic rules productively, using them in novel situations? We draw from the literature on determiner/noun productivity by training an unsupervised, autoencoder network measuring its ability to combine nouns with determiners. Our simple autoencoder creates combinations it has not previously encountered and produces a degree of overlap matching adults. While this preliminary work does not provide conclusive evidence for productivity, it warrants further investigation with more complex models. Further, this work helps lay the foundations for future collaboration between the deep learning and cognitive science communities.