Thomas Foster

Michael Matthews, Matthew Jackson, Michael Beukman et al.

A goal-conditioned reinforcement learning agent exploring an environment will see a wealth of information throughout a trajectory, most of which is discarded when only performing on-policy updates with respect to the commanded goal. All-goals learning, where each transition is used for learning off-policy with respect to every goal, allows agents to extract maximal information, however it is usually computationally infeasible when done via naive relabelling. This can be overcome by jointly outputting values and actions for every goal at once, allowing for efficient, parallel all-goals updates with a single pass through the network, in a process we call Learning Everything all at Once (LEO). We show that this approach significantly outperforms other methods on goal-conditioned Craftax and is competitive with existing baselines on continuous control environments, while achieving a >250x speed-up compared to all-goals relabelling. We then go on to show that this approach can be made even more powerful by using LEO as a teacher network, rather than a direct actor. We hope that, by unlocking all-goals learning at scale, LEO can serve as a useful tool for RL practitioners in complex environments. We open source our code.

Willem Röpke, Samuel Coward, Andrei Lupu et al.

Recent advances in reasoning models have yielded impressive results in mathematics and coding. However, most approaches rely on static datasets, which have been suggested to encourage memorisation and limit generalisation. We introduce DéjàQ, a framework that departs from this paradigm by jointly evolving a diverse set of synthetic mathematical problems alongside model training. This evolutionary process adapts to the model's ability throughout training, optimising problems for learnability. We propose two LLM-driven mutation strategies in which the model itself mutates the training data, either by altering contextual details or by directly modifying problem structure. We find that the model can generate novel and meaningful problems, and that these LLM-driven mutations improve RL training. We analyse key aspects of DéjàQ, including the validity of generated problems and computational overhead. Our results underscore the potential of dynamically evolving training data to enhance mathematical reasoning and indicate broader applicability, which we will support by open-sourcing our code.

Andrew M. Bean, Ryan Othniel Kearns, Angelika Romanou et al.

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as 'safety' and 'robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

William Lugoloobi, Thomas Foster, William Bankes et al.

Running LLMs with extended reasoning on every problem is expensive, but determining which inputs actually require additional compute remains challenging. We investigate whether their own likelihood of success is recoverable from their internal representations before generation, and if this signal can guide more efficient inference. We train linear probes on pre-generation activations to predict policy-specific success on math and coding tasks, substantially outperforming surface features such as question length and TF-IDF. Using E2H-AMC, which provides both human and model performance on identical problems, we show that models encode a model-specific notion of difficulty that is distinct from human difficulty, and that this distinction increases with extended reasoning. Leveraging these probes, we demonstrate that routing queries across a pool of models can exceed the best-performing model whilst reducing inference cost by up to 70\% on MATH, showing that internal representations enable practical efficiency gains even when they diverge from human intuitions about difficulty. Our code is available at: https://github.com/KabakaWilliam/llms_know_difficulty

Thomas Foster, Ioana Croitoru, Robert Dorfman et al.

In this work, we address in-context learning (ICL) for the task of image segmentation, introducing a novel approach that adapts a modern Video Object Segmentation (VOS) technique for visual in-context learning. This adaptation is inspired by the VOS method's ability to efficiently and flexibly learn objects from a few examples. Through evaluations across a range of support set sizes and on diverse segmentation datasets, our method consistently surpasses existing techniques. Notably, it excels with data containing classes not encountered during training. Additionally, we propose a technique for support set selection, which involves choosing the most relevant images to include in this set. By employing support set selection, the performance increases for all tested methods without the need for additional training or prompt tuning. The code can be found at https://github.com/v7labs/XMem_ICL/.

Thomas Foster, Anya Sims, Johannes Forkel et al.

Reinforcement learning is now widely adopted as the final stage of large language model training, especially for reasoning-style tasks such as maths problems. Typically, models attempt each question many times during a single training step and attempt to learn from their successes and failures. However, we demonstrate that throughout training with two popular algorithms (PPO and VinePPO) on two widely used datasets, many questions are either solved by all attempts - meaning they are already learned - or by none - providing no meaningful training signal. To address this, we adapt a method from the reinforcement learning literature - sampling for learnability - and apply it to the reinforcement learning stage of LLM training. Our curriculum prioritises questions with high variance of success, i.e. those where the agent sometimes succeeds, but not always. Our findings demonstrate that this curriculum consistently boosts training performance across multiple algorithms and datasets, paving the way for more efficient and effective reinforcement learning with LLMs.

Dulhan Jayalath, Shashwat Goel, Thomas Foster et al. · allen-ai

Where do learning signals come from when there is no ground truth in post-training? We propose turning exploration into supervision through Compute as Teacher (CaT), which converts the model's own exploration at inference-time into reference-free supervision by synthesizing a single reference from a group of parallel rollouts and then optimizing toward it. Concretely, the current policy produces a group of rollouts; a frozen anchor (the initial policy) reconciles omissions and contradictions to estimate a reference, turning extra inference-time compute into a teacher signal. We turn this into rewards in two regimes: (i) verifiable tasks use programmatic equivalence on final answers; (ii) non-verifiable tasks use self-proposed rubrics-binary, auditable criteria scored by an independent LLM judge, with reward given by the fraction satisfied. Unlike selection methods (best-of-N, majority, perplexity, or judge scores), synthesis may disagree with the majority and be correct even when all rollouts are wrong; performance scales with the number of rollouts. As a test-time procedure, CaT improves Gemma 3 4B, Qwen 3 4B, and Llama 3.1 8B (up to +27% on MATH-500; +12% on HealthBench). With reinforcement learning (CaT-RL), we obtain further gains (up to +33% and +30%), with the trained policy surpassing the initial teacher signal.

Bingchen Zhao, Despoina Magka, Minqi Jiang et al. · meta-ai, oxford

Rapid advancements in large language models (LLMs) have the potential to assist in scientific progress. A critical capability toward this endeavor is the ability to reproduce existing work. To evaluate the ability of AI agents to reproduce results in an active research area, we introduce the Automated LLM Speedrunning Benchmark, leveraging the research community contributions on the NanoGPT speedrun, a competition to train a GPT-2 model in the shortest time. Each of the 19 speedrun tasks provides the agent with the previous records training script, optionally paired with one of three hint formats, ranging from pseudocode to paper-like descriptions of the new records improvements. Records execute quickly by design and speedrun improvements encompass diverse code-level changes, ranging from high-level algorithmic advancements to hardware-aware optimizations. These features make the benchmark both accessible and realistic for the frontier problem of improving LLM training. We find that recent reasoning LLMs combined with SoTA scaffolds struggle to reimplement already-known innovations in our benchmark, even when given detailed hints. Our benchmark thus provides a simple, non-saturated measure of an LLMs ability to automate scientific reproduction, a necessary (but not sufficient) skill for an autonomous research agent.

Thomas Foster, Chon Lok Lei, Martin Robinson et al.

High dimensional integration is essential to many areas of science, ranging from particle physics to Bayesian inference. Approximating these integrals is hard, due in part to the difficulty of locating and sampling from regions of the integration domain that make significant contributions to the overall integral. Here, we present a new algorithm called Tree Quadrature (TQ) that separates this sampling problem from the problem of using those samples to produce an approximation of the integral. TQ places no qualifications on how the samples provided to it are obtained, allowing it to use state-of-the-art sampling algorithms that are largely ignored by existing integration algorithms. Given a set of samples, TQ constructs a surrogate model of the integrand in the form of a regression tree, with a structure optimised to maximise integral precision. The tree divides the integration domain into smaller containers, which are individually integrated and aggregated to estimate the overall integral. Any method can be used to integrate each individual container, so existing integration methods, like Bayesian Monte Carlo, can be combined with TQ to boost their performance. On a set of benchmark problems, we show that TQ provides accurate approximations to integrals in up to 15 dimensions; and in dimensions 4 and above, it outperforms simple Monte Carlo and the popular Vegas method.