Edoardo Cetin

Edoardo Cetin, Benjamin Chamberlain, Michael Bronstein et al.

We propose a new class of deep reinforcement learning (RL) algorithms that model latent representations in hyperbolic space. Sequential decision-making requires reasoning about the possible future consequences of current behavior. Consequently, capturing the relationship between key evolving features for a given task is conducive to recovering effective policies. To this end, hyperbolic geometry provides deep RL models with a natural basis to precisely encode this inherently hierarchical information. However, applying existing methodologies from the hyperbolic deep learning literature leads to fatal optimization instabilities due to the non-stationarity and variance characterizing RL gradient estimators. Hence, we design a new general method that counteracts such optimization challenges and enables stable end-to-end learning with deep hyperbolic representations. We empirically validate our framework by applying it to popular on-policy and off-policy RL algorithms on the Procgen and Atari 100K benchmarks, attaining near universal performance and generalization benefits. Given its natural fit, we hope future RL research will consider hyperbolic representations as a standard tool.

Edoardo Cetin, Stefano Peluchetti, Emilio Castillo et al.

Scaling autoregressive large language models (LLMs) has driven unprecedented progress but comes with vast computational costs. In this work, we tackle these costs by leveraging unstructured sparsity within an LLM's feedforward layers, the components accounting for most of the model parameters and execution FLOPs. To achieve this, we introduce a new sparse packing format and a set of CUDA kernels designed to seamlessly integrate with the optimized execution pipelines of modern GPUs, enabling efficient sparse computation during LLM inference and training. To substantiate our gains, we provide a quantitative study of LLM sparsity, demonstrating that simple L1 regularization can induce over 99% sparsity with negligible impact on downstream performance. When paired with our kernels, we show that these sparsity levels translate into substantial throughput, energy efficiency, and memory usage benefits that increase with model scale. We will release all code and kernels under an open-source license to promote adoption and accelerate research toward establishing sparsity as a practical axis for improving the efficiency and scalability of modern foundation models.

Edoardo Cetin, Philip J. Ball, Steve Roberts et al.

Off-policy reinforcement learning (RL) from pixel observations is notoriously unstable. As a result, many successful algorithms must combine different domain-specific practices and auxiliary losses to learn meaningful behaviors in complex environments. In this work, we provide novel analysis demonstrating that these instabilities arise from performing temporal-difference learning with a convolutional encoder and low-magnitude rewards. We show that this new visual deadly triad causes unstable training and premature convergence to degenerate solutions, a phenomenon we name catastrophic self-overfitting. Based on our analysis, we propose A-LIX, a method providing adaptive regularization to the encoder's gradients that explicitly prevents the occurrence of catastrophic self-overfitting using a dual objective. By applying A-LIX, we significantly outperform the prior state-of-the-art on the DeepMind Control and Atari 100k benchmarks without any data augmentation or auxiliary losses.

Tim Merino, Sam Earle, Ryunosuke Iwai et al.

We introduce Dream-Cubed, a large-scale dataset of Minecraft worlds at voxel resolution, and a family of models using cubes as powerful compositional units for efficient generation of interactive 3D environments. Dream-Cubed comprises tens of billions of tokens from a carefully curated mixture of procedural biome terrain and high-quality human-authored maps. We use this dataset to conduct the first large-scale study of 3D diffusion models for voxel generation, analyzing discrete and continuous diffusion formulations, data compositions, and architectural design choices. Our models operate directly in the space of blocks, enabling efficient and semantically grounded generation while supporting interactive user workflows such as inpainting and outpainting from user-authored blocks. To quantitatively evaluate our models, we adapt the FID metric to assess semantic differences between real and generated world renderings, and validate generation quality through a human preference study. We release the full dataset, code, and all our pretrained models, which we hope will provide a foundation for future research in efficient generative modeling for structured, interactive 3D environments.

Edoardo Cetin, Oya Celiktutan

We introduce a new framework that performs decision-making in reinforcement learning (RL) as an iterative reasoning process. We model agent behavior as the steady-state distribution of a parameterized reasoning Markov chain (RMC), optimized with a new tractable estimate of the policy gradient. We perform action selection by simulating the RMC for enough reasoning steps to approach its steady-state distribution. We show our framework has several useful properties that are inherently missing from traditional RL. For instance, it allows agent behavior to approximate any continuous distribution over actions by parameterizing the RMC with a simple Gaussian transition function. Moreover, the number of reasoning steps to reach convergence can scale adaptively with the difficulty of each action selection decision and can be accelerated by re-using past solutions. Our resulting algorithm achieves state-of-the-art performance in popular Mujoco and DeepMind Control benchmarks, both for proprioceptive and pixel-based tasks.

Robert Tjarko Lange, Yuki Imajuku, Edoardo Cetin

We introduce ShinkaEvolve: a new open-source framework leveraging large language models (LLMs) to advance scientific discovery with state-of-the-art performance and unprecedented efficiency. Recent advances in scaling inference time compute of LLMs have enabled significant progress in generalized scientific discovery. These approaches rely on evolutionary agentic harnesses that leverage LLMs as mutation operators to generate candidate solutions. However, current code evolution methods suffer from critical limitations: they are sample inefficient, requiring thousands of samples to identify effective solutions, and remain closed-source, hindering broad adoption and extension. ShinkaEvolve addresses these limitations, introducing three key innovations: a parent sampling technique balancing exploration and exploitation, code novelty rejection-sampling for efficient search space exploration, and a bandit-based LLM ensemble selection strategy. We evaluate ShinkaEvolve across diverse tasks, demonstrating consistent improvements in sample efficiency and solution quality. ShinkaEvolve discovers a new state-of-the-art circle packing solution using only 150 samples, designs high-performing agentic harnesses for AIME mathematical reasoning tasks, identifies improvements to ALE-Bench competitive programming solutions, and discovers novel mixture-of-expert load balancing loss functions that illuminate the space of optimization strategies. Our results demonstrate that ShinkaEvolve achieves broad applicability with exceptional sample efficiency. By providing open-source accessibility and cost-efficiency, this work democratizes open-ended discovery across diverse computational problems.

Edoardo Cetin, Tianyu Zhao, Yujin Tang

Training reasoning language models (LMs) with reinforcement learning (RL) for one-hot correctness inherently relies on the LM being able to explore and solve its task with some chance at initialization. Furthermore, a key use case of reasoning LMs is to act as teachers for distilling new students and cold-starting future RL iterations rather than being deployed themselves. From these considerations, we introduce a new framework that avoids RL's exploration challenge by training a new class of Reinforcement-Learned Teachers (RLTs) focused on yielding the most effective downstream distillation. RLTs are prompted with both the question and solution to each problem, and tasked to simply "connect-the-dots" with detailed explanations tailored for their students. We train RLTs with dense rewards obtained by feeding each explanation to the student and testing its understanding of the problem's solution. In practice, the raw outputs of a 7B RLT provide higher final performance on competition and graduate-level tasks than existing distillation and cold-starting pipelines that collect and postprocess the reasoning traces of orders of magnitude larger LMs. Furthermore, RLTs maintain their effectiveness when training larger students and when applied zero-shot to out-of-distribution tasks, unlocking new levels of efficiency and re-usability for the RL reasoning framework. Code available at: https://github.com/SakanaAI/RLT

Rujikorn Charakorn, Edoardo Cetin, Shinnosuke Uesaka et al.

Long input sequences are central to in-context learning, document understanding, and multi-step reasoning of Large Language Models (LLMs). However, the quadratic attention cost of Transformers makes inference memory-intensive and slow. While context distillation (CD) can transfer information into model parameters, per-prompt distillation is impractical due to training costs and latency. To address these limitations, we propose Doc-to-LoRA (D2L), a lightweight hypernetwork that meta-learns to perform approximate CD within a single forward pass. Given an unseen prompt, D2L generates a LoRA adapter for a target LLM, enabling subsequent queries to be answered without re-consuming the original context, reducing latency and KV-cache memory consumption during inference of the target LLM. On a long-context needle-in-a-haystack task, D2L successfully learns to map contexts into adapters that store the needle information, achieving near-perfect zero-shot accuracy at sequence lengths exceeding the target LLM's native context window by more than 4x. On real-world QA datasets with limited compute, D2L outperforms standard CD while significantly reducing peak memory consumption and update latency. We envision that D2L can facilitate rapid adaptation of LLMs, opening up the possibility of frequent knowledge updates and personalized chat behavior.

Stefan Nielsen, Edoardo Cetin, Peter Schwendeman et al.

Powerful large language models (LLMs) from different providers have been expensively trained and finetuned to specialize across varying domains. In this work, we introduce a new kind of Conductor model trained with reinforcement learning to automatically discover powerful coordination strategies among LLMs. Our Conductor learns not only to design targeted communication topologies for effective agent-to-agent collaboration, but also to prompt engineer focused instructions to the LLMs to maximally leverage their individual capabilities. We show that, by learning optimal coordination strategies over pools of powerful worker LLMs, a 7B Conductor achieves significant performance gains beyond any individual worker, attaining state-of-the-art results in challenging reasoning benchmarks, such as LiveCodeBench and GPQA. By training with randomized agent pools, our conductor effectively adapts to arbitrary sets of open- and closed-source agents, meeting any user requirements. Furthermore, allowing the Conductor to select itself as a worker gives rise to recursive topologies, elevating performance with a new form of dynamic test-time scaling through online iterative adaptation. More broadly, ours is among the early work demonstrating language model coordination can be unlocked through RL, where powerful coordination strategies emerge naturally in LLMs through pure end-to-end reward maximization.

Gianluigi Silvestri, Edoardo Cetin

Reasoning-oriented language models achieve strong performance by generating long chain-of-thought traces at inference time. However, this capability comes with substantial and often excessive computational cost, which can materialize in redundant or inefficient reasoning. We study this setting and introduce Truncated-Reasoning Self-Distillation (TRSD), a lightweight post-training procedure that encourages models to produce correct predictions from partial reasoning traces. In TRSD, a frozen teacher model first generates a full reasoning trace and evaluates the corresponding answer distribution conditioned on the prompt and the complete reasoning to construct a synthetic training target. A student model with the same architecture is then trained to match the teacher's answer distribution while being conditioned only on a truncated prefix of its reasoning trace. Across multiple reasoning benchmarks and token budgets, we demonstrate that TRSD improves robustness to truncated inference, with far reduced accuracy tradeoffs when applied to a diverse set of reasoning models. Moreover, although never explicitly regularized for shorter generation during training, we also find that TRSD-trained models inherently output shorter reasoning traces without truncation, significantly reducing inference-time costs even without artificial interventions.

Jinglue Xu, Qi Sun, Peter Schwendeman et al.

Combining diverse foundation models is promising, but weight-merging is limited by mismatched architectures and closed APIs. Trinity addresses this with a lightweight coordinator that orchestrates collaboration among large language models (LLMs). The coordinator, comprising a compact language model (approximately $0.6$B parameters) and a lightweight head (approximately $10$K parameters), is optimized with an evolutionary strategy for efficient and adaptive delegation. Trinity processes queries over multiple turns, where at each turn the coordinator assigns one of three roles (Thinker, Worker, or Verifier) to a selected LLM, effectively offloading complex skill acquisition from the coordinator itself. Experiments show that Trinity consistently outperforms individual models and existing methods across coding, math, reasoning, and domain knowledge tasks, and generalizes robustly to out-of-distribution tasks. On standard benchmarks, Trinity achieves state-of-the-art results, including a score of 86.2% on LiveCodeBench. Theoretical and empirical analyses identify two main factors behind this performance: (1) the coordinator's hidden-state representations provide rich contextualization of inputs, and (2) under high dimensionality and strict budget constraints, the separable Covariance Matrix Adaptation Evolution Strategy offers advantages over reinforcement learning, imitation learning, and random search by exploiting potential block-epsilon-separability.

Qi Sun, Edoardo Cetin, Yujin Tang

Self-adaptive large language models (LLMs) aim to solve the challenges posed by traditional fine-tuning methods, which are often computationally intensive and static in their ability to handle diverse tasks. We introduce Transformer-Squared, a novel self-adaptation framework that adapts LLMs for unseen tasks in real-time by selectively adjusting only the singular components of their weight matrices. During inference, Transformer-Squared employs a two-pass mechanism: first, a dispatch system identifies the task properties, and then task-specific 'expert' vectors, trained using reinforcement learning, are dynamically mixed to obtain targeted behavior for the incoming prompt. Our method consistently outperforms ubiquitous approaches such as LoRA, with fewer parameters and greater efficiency. Furthermore, Transformer-Squared demonstrates versatility across different LLM architectures and modalities, including vision-language tasks. Transformer-Squared represents a significant leap forward, offering a scalable, efficient solution for enhancing the adaptability and task-specific performance of LLMs, paving the way for truly dynamic, self-organizing AI systems.

Jeffrey Seely, Yuki Imajuku, Tianyu Zhao et al.

Existing reasoning benchmarks for large language models (LLMs) frequently fail to capture authentic creativity, often rewarding memorization of previously observed patterns. We address this shortcoming with Sudoku-Bench, a curated benchmark of challenging and unconventional Sudoku variants specifically selected to evaluate creative, multi-step logical reasoning. Sudoku variants form an unusually effective domain for reasoning research: each puzzle introduces unique or subtly interacting constraints, making memorization infeasible and requiring solvers to identify novel logical breakthroughs (``break-ins''). Despite their diversity, Sudoku variants maintain a common and compact structure, enabling clear and consistent evaluation. Sudoku-Bench includes a carefully chosen puzzle set, a standardized text-based puzzle representation, and flexible tools compatible with thousands of publicly available puzzles -- making it easy to extend into a general research environment. Baseline experiments show that state-of-the-art LLMs solve fewer than 15\% of puzzles unaided, highlighting significant opportunities to advance long-horizon, strategic reasoning capabilities.

Edoardo Cetin, Ahmed Touati, Yann Ollivier

The forward-backward representation (FB) is a recently proposed framework (Touati et al., 2023; Touati & Ollivier, 2021) to train behavior foundation models (BFMs) that aim at providing zero-shot efficient policies for any new task specified in a given reinforcement learning (RL) environment, without training for each new task. Here we address two core limitations of FB model training. First, FB, like all successor-feature-based methods, relies on a linear encoding of tasks: at test time, each new reward function is linearly projected onto a fixed set of pre-trained features. This limits expressivity as well as precision of the task representation. We break the linearity limitation by introducing auto-regressive features for FB, which let finegrained task features depend on coarser-grained task information. This can represent arbitrary nonlinear task encodings, thus significantly increasing expressivity of the FB framework. Second, it is well-known that training RL agents from offline datasets often requires specific techniques.We show that FB works well together with such offline RL techniques, by adapting techniques from (Nair et al.,2020b; Cetin et al., 2024) for FB. This is necessary to get non-flatlining performance in some datasets, such as DMC Humanoid. As a result, we produce efficient FB BFMs for a number of new environments. Notably, in the D4RL locomotion benchmark, the generic FB agent matches the performance of standard single-task offline agents (IQL, XQL). In many setups, the offline techniques are needed to get any decent performance at all. The auto-regressive features have a positive but moderate impact, concentrated on tasks requiring spatial precision and task generalization beyond the behaviors represented in the trainset.

Edoardo Cetin, Tianyu Zhao, Yujin Tang

We propose a new finetuning method to provide pre-trained large language models (LMs) the ability to scale test-time compute through the diffusion framework. By increasing the number of diffusion steps, we show our finetuned models achieve monotonically increasing accuracy, directly translating to improved performance across downstream tasks. Furthermore, our finetuned models can expertly answer questions on specific topics by integrating powerful guidance techniques, and autonomously determine the compute required for a given problem by leveraging adaptive ODE solvers. Our method is universally applicable to any foundation model pre-trained with a cross-entropy loss and does not modify any of its original weights, fully preserving its strong single-step generation capabilities. We show our method is more effective and fully compatible with traditional finetuning approaches, introducing an orthogonal new direction to unify the strengths of the autoregressive and diffusion frameworks.

Edoardo Cetin, Qi Sun, Tianyu Zhao et al.

Prior methods propose to offset the escalating costs of modern foundation models by dropping specific parts of their contexts with hand-designed rules, while attempting to preserve their original performance. We overcome this trade-off with Neural Attention Memory Models (NAMMs), introducing a learned network for memory management that improves both the performance and efficiency of transformers. We evolve NAMMs atop pre-trained transformers to provide different latent contexts focusing on the most relevant information for individual layers and attention heads. NAMMs are universally applicable to any model using self-attention as they condition exclusively on the values in the produced attention matrices. Learning NAMMs on a small set of problems, we achieve substantial performance improvements across multiple long-context benchmarks while cutting the model's input contexts up to a fraction of the original sizes. We show the generality of our conditioning enables zero-shot transfer of NAMMs trained only on language to entirely new transformer architectures even across input modalities, with their benefits carrying over to vision and reinforcement learning.

Yoav Gelberg, Koshi Eguchi, Takuya Akiba et al.

So far, expensive finetuning beyond the pretraining sequence length has been a requirement for effectively extending the context of language models (LM). In this work, we break this key bottleneck by Dropping the Positional Embeddings of LMs after training (DroPE). Our simple method is motivated by three key theoretical and empirical observations. First, positional embeddings (PEs) serve a crucial role during pretraining, providing an important inductive bias that significantly facilitates convergence. Second, over-reliance on this explicit positional information is also precisely what prevents test-time generalization to sequences of unseen length, even when using popular PE-scaling methods. Third, positional embeddings are not an inherent requirement of effective language modeling and can be safely removed after pretraining, following a short recalibration phase. Empirically, DroPE yields seamless zero-shot context extension without any long-context finetuning, quickly adapting pretrained LMs without compromising their capabilities in the original training context. Our findings hold across different models and dataset sizes, far outperforming previous specialized architectures and established rotary positional embedding scaling methods.

Edoardo Cetin, Andrea Tirinzoni, Matteo Pirotta et al.

Offline reinforcement learning algorithms have proven effective on datasets highly connected to the target downstream task. Yet, leveraging a novel testbed (MOOD) in which trajectories come from heterogeneous sources, we show that existing methods struggle with diverse data: their performance considerably deteriorates as data collected for related but different tasks is simply added to the offline buffer. In light of this finding, we conduct a large empirical study where we formulate and test several hypotheses to explain this failure. Surprisingly, we find that scale, more than algorithmic considerations, is the key factor influencing performance. We show that simple methods like AWAC and IQL with increased network size overcome the paradoxical failure modes from the inclusion of additional data in MOOD, and notably outperform prior state-of-the-art algorithms on the canonical D4RL benchmark.

Edoardo Cetin, Oya Celiktutan

Popular off-policy deep reinforcement learning algorithms compensate for overestimation bias during temporal-difference learning by utilizing pessimistic estimates of the expected target returns. In this work, we propose a novel learnable penalty to enact such pessimism, based on a new way to quantify the critic's epistemic uncertainty. Furthermore, we propose to learn the penalty alongside the critic with dual TD-learning, a strategy to estimate and minimize the bias magnitude in the target returns. Our method enables us to accurately counteract overestimation bias throughout training without incurring the downsides of overly pessimistic targets. Empirically, by integrating our method and other orthogonal improvements with popular off-policy algorithms, we achieve state-of-the-art results in continuous control tasks from both proprioceptive and pixel observations.

Edoardo Cetin, Oya Celiktutan

The performance of reinforcement learning depends upon designing an appropriate action space, where the effect of each action is measurable, yet, granular enough to permit flexible behavior. So far, this process involved non-trivial user choices in terms of the available actions and their execution frequency. We propose a novel framework for reinforcement learning that effectively lifts such constraints. Within our framework, agents learn effective behavior over a routine space: a new, higher-level action space, where each routine represents a set of 'equivalent' sequences of granular actions with arbitrary length. Our routine space is learned end-to-end to facilitate the accomplishment of underlying off-policy reinforcement learning objectives. We apply our framework to two state-of-the-art off-policy algorithms and show that the resulting agents obtain relevant performance improvements while requiring fewer interactions with the environment per episode, improving computational efficiency.

Jian Jiang, Edoardo Cetin, Oya Celiktutan

Incremental learning aims to enable machine learning models to continuously acquire new knowledge given new classes, while maintaining the knowledge already learned for old classes. Saving a subset of training samples of previously seen classes in the memory and replaying them during new training phases is proven to be an efficient and effective way to fulfil this aim. It is evident that the larger number of exemplars the model inherits the better performance it can achieve. However, finding a trade-off between the model performance and the number of samples to save for each class is still an open problem for replay-based incremental learning and is increasingly desirable for real-life applications. In this paper, we approach this open problem by tapping into a two-step compression approach. The first step is a lossy compression, we propose to encode input images and save their discrete latent representations in the form of codes that are learned using a hierarchical Vector Quantised Variational Autoencoder (VQ-VAE). In the second step, we further compress codes losslessly by learning a hierarchical latent variable model with bits-back asymmetric numeral systems (BB-ANS). To compensate for the information lost in the first step compression, we introduce an Information Back (IB) mechanism that utilizes real exemplars for a contrastive learning loss to regularize the training of a classifier. By maintaining all seen exemplars' representations in the format of `codes', Discrete Representation Replay (DRR) outperforms the state-of-art method on CIFAR-100 by a margin of 4% accuracy with a much less memory cost required for saving samples. Incorporated with IB and saving a small set of old raw exemplars as well, the accuracy of DRR can be further improved by 2% accuracy.

Edoardo Cetin, Oya Celiktutan

Human beings are able to understand objectives and learn by simply observing others perform a task. Imitation learning methods aim to replicate such capabilities, however, they generally depend on access to a full set of optimal states and actions taken with the agent's actuators and from the agent's point of view. In this paper, we introduce a new algorithm - called Disentangling Generative Adversarial Imitation Learning (DisentanGAIL) - with the purpose of bypassing such constraints. Our algorithm enables autonomous agents to learn directly from high dimensional observations of an expert performing a task, by making use of adversarial learning with a latent representation inside the discriminator network. Such latent representation is regularized through mutual information constraints to incentivize learning only features that encode information about the completion levels of the task being demonstrated. This allows to obtain a shared feature space to successfully perform imitation while disregarding the differences between the expert's and the agent's domains. Empirically, our algorithm is able to efficiently imitate in a diverse range of control problems including balancing, manipulation and locomotive tasks, while being robust to various domain differences in terms of both environment appearance and agent embodiment.