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.

As language models are increasingly deployed as autonomous agents, they must coordinate with others over long horizons in open-ended interactive tasks. Yet existing evaluations rarely test these demands together, instead emphasising single-agent tasks, short interactions, or highly structured multi-agent settings. We introduce alem, a JAX-based benchmark for open-ended multi-agent coordination built on Craftax-like dynamics. Alem embeds procedurally generated coordination tasks, soft specialisation, communication, and controllable coordination difficulty into a long-horizon survival world with exploration, crafting, trading, and combat. We evaluate 13 modern LLMs zero-shot within homogeneous teams, with trained MARL agents as reference points. Current LLM agents remain far from solving alem, averaging only 6

Much recent progress in Large Language Model (LLM) performance has been driven by verifiable feedback in deterministic domains like mathematics and code. However, scaling reinforcement learning (RL) and test-time compute in domains for which strict verification is infeasible remains a challenge. A common approach is to use an LLM-as-judge, which often relies on opaque, monolithic scores. In this work, we propose that AI feedback is most effective when decomposed into granular, prompt-specific checklists. To transform these checklists into a scalar reward, we introduce DIVA: DIscriminative VAriance weighting, a dynamic aggregation scheme that prioritises checklist items based on their ability to distinguish quality across a candidate pool. This ensures the reward signal focuses on the most salient criteria for a given prompt and response group, rather than being diluted by trivial or redundant constraints. Our approach yields an 11.8% win-rate improvement on AlpacaEval 2.0 using Qwen3-8B, outperforming holistic reward models and existing checklist baselines. Beyond training, we show that these checklists serve as a structured policy improvement operator at inference time - by using the model’s own checklist evaluation as localised contextual feedback, the model can iteratively refine its output. This self-correction mechanism outperforms free-form sequential self-correction, offering a unified and interpretable framework for scaling both training-time and test-time performance in domains lacking strict verifiers.

Large language models (LLMs) are typically trained to acquire behaviours from demonstrations or experience, yet much of their training data is declarative: instructions, rules, and descriptions that specify behaviours without showing how to execute them. We introduce Programming by Backprop (PBB): a training regime that enables LLMs to acquire procedural knowledge (i.e., reusable behaviours) from declarative instructions encountered during training. With PBB, instructions in training data provide an opportunity to `program' specific behaviours into model weights. The core principle underpinning PBB is the separation of learning how instructions map to behaviour from internalising new instructions. We devise two distinct PBB curricula that leverage this principle. Through controlled experiments across two domains (algorithmic execution from Python source code and text generation from context-free grammars), we demonstrate the benefit of these curricula over training on a homogeneous data mixture. Crucially, PBB is highly sample efficient, with a single instruction substituting for up to 100 execution examples. Though execution of instructions in training data remains less reliable than when instructions are given in-context, our results demonstrate that procedural knowledge can be noisily `programmed' into LLMs through PBB, with important implications for data curation and safety.

Large Language Models (LLMs) have demonstrated remarkable improvements in reasoning and planning through increased test-time compute, often by framing problem-solving as a search process. While methods like Monte Carlo Tree Search (MCTS) have proven effective in some domains, their reliance on fixed exploration hyperparameters limits their adaptability across tasks of varying difficulty, rendering them impractical or expensive in certain settings. In this paper, we propose \textbf{LLM-First Search (LFS)}, a novel \textit{LLM Self-Guided Search} method that removes the need for pre-defined search strategies by empowering the LLM to autonomously control the search process via self-guided exploration. Rather than relying on external heuristics or hardcoded policies, the LLM evaluates whether to pursue the current search path or explore alternative branches based on its internal scoring mechanisms. This enables more flexible and context-sensitive reasoning without requiring manual tuning or task-specific adaptation. We evaluate LFS on Countdown and Sudoku against three classic widely-used search algorithms, Tree-of-Thoughts' Breadth First Search (ToT-BFS), Best First Search (BestFS), and MCTS, each of which have been used to achieve SotA results on a range of challenging reasoning tasks. We found that LFS (1) performs better on more challenging tasks without additional tuning, (2) is more computationally efficient compared to the other methods, especially when powered by a stronger model, (3) scales better with stronger models, due to its LLM-First design, and (4) scales better with increased compute budget. Our code is publicly available at \href{https://github.com/NathanHerr/LLM-First-Search}{LLM-First-Search}.

Training agents to act in embodied environments typically requires vast training data or access to accurate simulation, neither of which exists for many cases in the real world. Instead, world models are emerging as an alternative–leveraging offline, passively collected data, they make it possible to generate diverse worlds for training agents in simulation. In this work, we harness world models to generate “imagined” environments to train robust agents capable of generalizing to novel task variations. One of the challenges in doing this is ensuring the agent trains on useful generated data. We thus propose a novel approach IMAC (Imagined Autocurricula) leveraging Unsupervised Environment Design (UED), induces an automatic curriculum over generated worlds. In a series of challenging, procedurally generated environments, we show it is possible to achieve strong transfer performance on held-out environments having trained only inside a world model learned from a narrower dataset. We believe this opens the path to utilizing larger-scale, foundation world models for generally capable agents.

Automatic evaluation methods based on large language models (LLMs) are emerging as the standard tool for assessing the instruction-following abilities of LLM-based agents. The most common method in this paradigm, pairwise comparisons with a baseline model, critically depends on the assumption of transitive preferences. However, the validity of this assumption remains largely unexplored. In this study, we investigate the presence of non-transitivity within the AlpacaEval framework and analyze its effects on model rankings. We find that LLM judges exhibit non-transitive preferences, leading to rankings that are sensitive to the choice of the baseline model. To mitigate this issue, we show that round-robin tournaments combined with Bradley-Terry models of preference can produce more reliable rankings. Notably, our method increases both the Spearman correlation and the Kendall correlation with Chatbot Arena (95.0% -> 96.4% and 82.1% -> 86.3% respectively). To address the computational cost of round-robin tournaments, we propose Swiss-Wise Iterative Matchmaking (Swim) tournaments, using a dynamic matching strategy to capture the benefits of round-robin tournaments while maintaining computational efficiency.

Large Language Models (LLMs) and Vision Language Models (VLMs) possess extensive knowledge and exhibit promising reasoning abilities, however, they still struggle to perform well in complex, dynamic environments. Real-world tasks require handling intricate interactions, advanced spatial reasoning, long-term planning, and continuous exploration of new strategies—areas in which we lack effective methodologies for comprehensively evaluating these capabilities. To address this gap, we introduce BALROG, a novel benchmark designed to assess the agentic capabilities of LLMs and VLMs through a diverse set of challenging games. Our benchmark incorporates a range of existing reinforcement learning environments with varying levels of difficulty, including tasks that are solvable by non-expert humans in seconds to extremely challenging ones that may take years to master (e.g., the NetHack Learning Environment). We devise fine-grained metrics to measure performance and conduct an extensive evaluation of several popular open-source and closed-source LLMs and VLMs. Our findings indicate that while current models achieve partial success in the easier games, they struggle significantly with more challenging tasks. Notably, we observe severe deficiencies in vision-based decision-making, as several models perform worse when visual representations of the environments are provided. We release BALROG as an open and user-friendly benchmark to facilitate future research and development in the agentic community. Code and Leaderboard at balrogai.com

Training large language models (LLMs) to reason via reinforcement learning (RL) significantly improves their problem-solving capabilities. In agentic settings, existing methods like ReAct prompt LLMs to explicitly plan before every action; however, we demonstrate that always planning is computationally expensive and degrades performance on long-horizon tasks, while never planning further limits performance. To address this, we introduce a conceptual framework formalizing dynamic planning for LLM agents, enabling them to flexibly decide when to allocate test-time compute for planning. We propose a simple two-stage training pipeline: (1) supervised fine-tuning on diverse synthetic data to prime models for dynamic planning, and (2) RL to refine this capability in long-horizon environments. Experiments on the Crafter environment show that dynamic planning agents trained with this approach are more sample-efficient and consistently achieve more complex objectives. Additionally, we demonstrate that these agents can be effectively steered by human-written plans, surpassing their independent capabilities. To our knowledge, this work is the first to explore training LLM agents for dynamic test-time compute allocation in sequential decision-making tasks, paving the way for more efficient, adaptive, and controllable agentic systems.

Diffusion models have achieved state-of-the-art performance across multiple domains, with recent advancements extending their applicability to discrete data. However, aligning discrete diffusion models with task-specific preferences remains challenging, particularly in scenarios where explicit reward functions are unavailable. In this work, we introduce Discrete Diffusion DPO (D2-DPO), the first adaptation of Direct Preference Optimization (DPO) to discrete diffusion models formulated as continuous-time Markov chains. Our approach derives a novel loss function that directly fine-tunes the generative process using preference data while preserving fidelity to a reference distribution. We validate D2-DPO on a structured binary sequence generation task, demonstrating that the method effectively aligns model outputs with preferences while maintaining structural validity. Our results highlight that D2-DPO enables controlled fine-tuning without requiring explicit reward models, making it a practical alternative to reinforcement learning-based approaches. Future research will explore extending D2-DPO to more complex generative tasks, including language modeling and protein sequence generation, as well as investigating alternative noise schedules, such as uniform noising, to enhance flexibility across different applications.

The capabilities and limitations of Large Language Models have been sketched out in great detail in recent years, providing an intriguing yet conflicting picture. On the one hand, LLMs demonstrate a general ability to solve problems. On the other hand, they show surprising reasoning gaps when compared to humans, casting doubt on the robustness of their generalisation strategies. The sheer volume of data used in the design of LLMs has precluded us from applying the method traditionally used to measure generalisation: train-test set separation. To overcome this, we study what kind of generalisation strategies LLMs employ when performing reasoning tasks by investigating the pretraining data they rely on. For two models of different sizes (7B and 35B) and 2.5B of their pretraining tokens, we identify what documents influence the model outputs for three simple mathematical reasoning tasks and contrast this to the data that are influential for answering factual questions. We find that, while the models rely on mostly distinct sets of data for each factual question, a document often has a similar influence across different reasoning questions within the same task, indicating the presence of procedural knowledge. We further find that the answers to factual questions often show up in the most influential data. However, for reasoning questions the answers usually do not show up as highly influential, nor do the answers to the intermediate reasoning steps. When we characterise the top ranked documents for the reasoning questions qualitatively, we confirm that the influential documents often contain procedural knowledge, like demonstrating how to obtain a solution using formulae or code. Our findings indicate that the approach to reasoning the models use is unlike retrieval, and more like a generalisable strategy that synthesises procedural knowledge from documents doing a similar form of reasoning.

In recent years there has been a tremendous surge in the general capabilities of AI systems, mainly fuelled by training foundation models on internet-scale data. Nevertheless, the creation of open-ended, ever self-improving AI remains elusive.In this position paper, we argue that the ingredients are now in place to achieveopen-endednessin AI systems with respect to a human observer. Furthermore, we claim that such open-endedness is an essential property of any artificial superhuman intelligence (ASI).We begin by providing a concrete formal definition of open-endedness through the lens of novelty and learnability. We then illustrate a path towards ASI via open-ended systems built on top of foundation models, capable of making novel, human-relevant discoveries. We conclude by examining the safety implications of generally-capable open-ended AI. We expect that open-ended foundation models will prove to be an increasingly fertile and safety-critical area of research in the near future.

In multi-agent settings with mixed incentives, methods developed for zero-sum games have been shown to lead to detrimental outcomes. To address this issue, opponent shaping (OS) methods explicitly learn to influence the learning dynamics of co-players and empirically lead to improved individual and collective outcomes. However, OS methods have only been evaluated in low-dimensional environments due to the challenges associated with estimating higher-order derivatives or scaling model-free meta-learning. Alternative methods that scale to more complex settings either converge to undesirable solutions or rely on unrealistic assumptions about the environment or co-players. In this paper, we successfully scale an OS-based approach to general-sum games with temporally-extended actions and long-time horizons for the first time. After analysing the representations of the meta-state and history used by previous algorithms, we propose a simplified version called Shaper. We show empirically that Shaper leads to improved individual and collective outcomes in a range of challenging settings from literature. We further formalize a technique previously implicit in the literature, and analyse its contribution to opponent shaping. We show empirically that this technique is helpful for the functioning of prior methods in certain environments. Lastly, we show that previous environments, such as the CoinGame, are inadequate for analysing temporally-extended general-sum interactions.

Fine-tuning large pre-trained models has become the de facto strategy for developing both task-specific and general-purpose machine learning systems, including developing models that are safe to deploy. Despite its clear importance, there has been minimal work that explains how fine-tuning alters the underlying capabilities learned by a model during pretraining: does fine-tuning yield entirely novel capabilities or does it just modulate existing ones? We address this question empirically in synthetic, controlled settings where we can use mechanistic interpretability tools (e.g., network pruning and probing) to understand how the model's underlying capabilities are changing. We perform an extensive analysis of the effects of fine-tuning in these settings, and show that: (i) fine-tuning rarely alters the underlying model capabilities; (ii) a minimal transformation, which we call a 'wrapper', is typically learned on top of the underlying model capabilities, creating the illusion that they have been modified; and (iii) further fine-tuning on a task where such hidden capabilities are relevant leads to sample-efficient 'revival' of the capability, i.e., the model begins reusing these capability after only a few gradient steps. This indicates that practitioners can unintentionally remove a model's safety wrapper merely by fine-tuning it on a, e.g., superficially unrelated, downstream task. We additionally perform analysis on language models trained on the TinyStories dataset to support our claims in a more realistic setup.

Post-Training Quantization (PTQ) enhances the efficiency of Large Language Models (LLMs) by enabling faster operation and compatibility with more accessible hardware through reduced memory usage, at the cost of small performance drops. We explore the role of calibration sets in PTQ, specifically their effect on hidden activations in various notable open-source LLMs. Calibration sets are crucial for evaluating activation magnitudes and identifying outliers, which can distort the quantization range and negatively impact performance. Our analysis reveals a marked contrast in quantization effectiveness across models. The older OPT model, which much of the quantization literature is based on, shows significant performance deterioration and high susceptibility to outliers with varying calibration sets. In contrast, newer models like Llama-2 7B, Llama-3 8B, Command-R 35B, and Mistral 7B demonstrate strong robustness, with Mistral 7B showing near-immunity to outliers and stable activations. These findings suggest a shift in PTQ strategies might be needed. As advancements in pre-training methods reduce the relevance of outliers, there is an emerging need to reassess the fundamentals of current quantization literature. The emphasis should pivot towards optimizing inference speed, rather than primarily focusing on outlier preservation, to align with the evolving characteristics of state-of-the-art LLMs.

In the rapidly advancing field of multi-agent systems, ensuring robustness in unfamiliar and adversarial settings is crucial. Notwithstanding their outstanding performance in familiar environments, these systems often falter in new situations due to overfitting during the training phase. This is especially pronounced in settings where both cooperative and competitive behaviours are present, encapsulating a dual nature of overfitting and generalisation challenges. To address this issue, we present Multi-Agent Diagnostics for Robustness via Illuminated Diversity (MADRID), a novel approach for generating diverse adversarial scenarios that expose strategic vulnerabilities in pre-trained multi-agent policies. Leveraging the concepts from open-ended learning, MADRID navigates the vast space of adversarial settings, employing a target policy's regret to gauge the vulnerabilities of these settings. We evaluate the effectiveness of MADRID on the 11vs11 version of Google Research Football, one of the most complex environments for multi-agent reinforcement learning. Specifically, we employ MADRID for generating a diverse array of adversarial settings for TiZero, the state-of-the-art approach which "masters" the game through 45 days of training on a large-scale distributed infrastructure. We expose key shortcomings in TiZero's tactical decision-making, underlining the crucial importance of rigorous evaluation in multi-agent systems.

As large language models (LLMs) become increasingly prevalent across many real-world applications, understanding and enhancing their robustness to adversarial attacks is of paramount importance. Existing methods for identifying adversarial prompts tend to focus on specific domains, lack diversity, or require extensive human annotations. To address these limitations, we present Rainbow Teaming, a novel black-box approach for producing a diverse collection of adversarial prompts. Rainbow Teaming casts adversarial prompt generation as a quality-diversity problem and uses open-ended search to generate prompts that are both effective and diverse. Focusing on the safety domain, we use Rainbow Teaming to target various state-of-the-art LLMs, including the Llama 2 and Llama 3 models. Our approach reveals hundreds of effective adversarial prompts, with an attack success rate exceeding 90% across all tested models. Furthermore, we demonstrate that prompts generated by Rainbow Teaming are highly transferable and that fine-tuning models with synthetic data generated by our method significantly enhances their safety without sacrificing general performance or helpfulness. We additionally explore the versatility of Rainbow Teaming by applying it to question answering and cybersecurity, showcasing its potential to drive robust open-ended self-improvement in a wide range of applications.

Benchmarks play an important role in the development of machine learning algorithms, with reinforcement learning (RL) research having been heavily influenced by the available environments. However, RL environments are traditionally run on the CPU, limiting their scalability with typical academic compute. Recent advancements in JAX have enabled the wider use of hardware acceleration to overcome these computational hurdles, enabling massively parallel RL training pipelines and environments. This is particularly useful for multi-agent reinforcement learning (MARL) research. First of all, multiple agents must be considered at each environment step, adding computational burden, and secondly, the sample complexity is increased due to non-stationarity, decentralised partial observability, or other MARL challenges. In this paper, we present JaxMARL, the first open-source code base that combines ease-of-use with GPU enabled efficiency, and supports a large number of commonly used MARL environments as well as popular baseline algorithms. When considering wall clock time, our experiments show that per-run our JAX-based training pipeline is up to 12500x faster than existing approaches. %This enables efficient and thorough evaluations, with the potential to alleviate the evaluation crisis of the field. We also introduce and benchmark SMAX, a vectorised, simplified version of the popular StarCraft Multi-Agent Challenge, which removes the need to run the StarCraft II game engine. This not only enables GPU acceleration, but also provides a more flexible MARL environment, unlocking the potential for self-play, meta-learning, and other future applications in MARL. We provide code at https://github.com/flairox/jaxmarl.

Humanoid control is an important research challenge offering avenues for integration into human-centric infrastructures and enabling physics-driven humanoid animations. The daunting challenges in this field stem from the difficulty of optimizing in high-dimensional action spaces and the instability introduced by the bipedal morphology of humanoids. However, the extensive collection of human motion-captured data and the derived datasets of humanoid trajectories, such as MoCapAct, paves the way to tackle these challenges. In this context, we present Humanoid Generalist Autoencoding Planner (H-GAP), a state-action trajectory generative model trained on humanoid trajectories derived from human motion-captured data, capable of adeptly handling downstream control tasks with Model Predictive Control (MPC). For 56 degrees of freedom humanoid, we empirically demonstrate that H-GAP learns to represent and generate a wide range of motor behaviours. Further, without any learning from online interactions, it can also flexibly transfer these behaviors to solve novel downstream control tasks via planning. Notably, H-GAP excels established MPC baselines that have access to the ground truth dynamics model, and is superior or comparable to offline RL methods trained for individual tasks. Finally, we do a series of empirical studies on the scaling properties of H-GAP, showing the potential for performance gains via additional data but not computing. Code and videos are available at https://ycxuyingchen.github.io/hgap/.

Common methods for aligning large language models (LLMs) with desired behaviour heavily rely on human-labelled data. However, as models grow increasingly sophisticated, they will surpass human expertise, and the role of human evaluation will evolve into non-experts overseeing experts. In anticipation of this, we ask: can weaker models assess the correctness of stronger models? We investigate this question in an analogous setting, where stronger models (experts) possess the necessary information to answer questions and weaker models (non-experts) lack this information. The method we evaluate isdebate, where two LLM experts each argue for a different answer, and a non-expert selects the answer. We find that debate consistently helps both non-expert models and humans answer questions, achieving 76% and 88% accuracy respectively (naive baselines obtain 48% and 60%). Furthermore, optimising expert debaters for persuasiveness in an unsupervised manner improves non-expert ability to identify the truth in debates. Our results provide encouraging empirical evidence for the viability of aligning models with debate in the absence of ground truth.