Neural quantum states (NQS) provide a flexible and scalable framework for approximating quantum many-body wavefunctions. Among NQS parameterizations, autoregressive models are especially attractive because they enable exact, independent sampling from the Born distribution, avoiding the autocorrelation and mixing issues of Markov chain methods. Yet their optimization remains comparatively underexplored: Adam is a scalable method but ignores function space geometry, while stochastic reconfiguration is principled but costly and numerically fragile in large models. To address this gap, we show that variational energy minimization can be viewed as an advantage policy-gradient problem over the Born distribution, motivating trust-region optimization for NQS training. We introduce Proximal Wavefunction Optimization (PWO), a principled trust-region algorithm that clips probability-ratio changes in the amplitude channel and phase increments in the phase channel. PWO avoids explicit matrix inversion, reuses samples across multiple updates, and combines the scalability of first-order optimization with theoretical guarantees. Across Ising and frustrated J_1-J_2 one- and two-dimensional spin systems, PWO improves stability and wall-clock convergence over Adam, minSR, and SPRING. Finally, we fine-tune a 1.5B-parameter RWKV-7 model, demonstrating NQS optimization at a scale over three orders of magnitude beyond prior work.

The availability of large amounts of clean data is paramount to training neural networks. However, at large scales, manual oversight is impractical, resulting in sizeable datasets that can be very noisy. Attempts to mitigate this obstacle to producing performant vision-language models have so far involved heuristics, curated reference datasets, and using pre-trained models. Here we propose a novel, bootstrapped method in which a CLIP model is trained on an evolving, self-selected dataset. This evolving dataset constitutes a balance of filtered, highly probable clean samples as well as diverse samples from the entire distribution. Our proposed Self-Filtering method iterates between training the model and selecting a subsequently improved data mixture. Training on vision-language datasets filtered by the proposed approach improves downstream performance without the need for additional data or pre-trained models.

Text-to-image (T2I) models offer great potential for creating virtually limitless synthetic data, a valuable resource compared to fixed and finite real datasets. Previous works evaluate the utility of synthetic data from T2I models on three key desiderata: quality, diversity, and consistency. While prompt engineering is the primary means of interacting with T2I models, the systematic impact of prompt complexity on these critical utility axes remains underexplored. In this paper, we first conduct synthetic experiments to motivate the difficulty of generalization w.r.t. prompt complexity and explain the observed difficulty with theoretical derivations. Then, we introduce a new evaluation framework that can compare the utility of real data and synthetic data, and present a comprehensive analysis of how prompt complexity influences the utility of synthetic data generated by commonly used T2I models. We conduct our study across diverse datasets, including CC12M, ImageNet-1k, and DCI, and evaluate different inference-time intervention methods. Our synthetic experiments show that generalizing to more general conditions is harder than the other way round, since the former needs an estimated likelihood that is not learned by diffusion models. Our large-scale empirical experiments reveal that increasing prompt complexity results in lower conditional diversity and prompt consistency, while reducing the synthetic-to-real distribution shift, which aligns with the synthetic experiments. Moreover, current inference-time interventions can augment the diversity of the generations at the expense of moving outside the support of real data. Among those interventions, prompt expansion, by deliberately using a pre-trained language model as a likelihood estimator, consistently achieves the highest performance in both image diversity and aesthetics, even higher than that of real data. Combining advanced guidance interventions with prompt expansion results in the most appealing utility trade-offs of synthetic data.

Evolution Strategies (ES) is a class of powerful black-box optimisation methods that are highly parallelisable and can handle non-differentiable and noisy objectives. However, naïve ES becomes prohibitively expensive at scale on GPUs due to the low arithmetic intensity of batched matrix multiplications with unstructured random perturbations. We introduce Evolution Guided GeneRal Optimisation via Low-rank Learning (EGGROLL), which improves arithmetic intensity by structuring individual perturbations as rank-r matrices, resulting in a hundredfold increase in training speed for billion-parameter models at large population sizes, achieving up to 91

Reasoning LLMs suffer from quadratic compute growth as their context length increases, making reinforcement learning with verifiable rewards (RLVR) and test-time scaling prohibitively expensive. Prior work has tried to lighten the computational burden by shortening reasoning traces through pruning, summarization, or multi-stage training, but these methods remain bound to quadratic costs. We introduce Delethink, a thinking algorithm that realizes the Markovian Thinking Paradigm. Instead of producing one long monolithic reasoning trace, Delethink thinks in a sequence of chunks, the Delethink trace. Each chunk continues reasoning by referring only to a fixed number of prior tokens, which functions as a Markovian state sufficient for progressing reasoning, while deleting the rest. This preserves continuity without carrying the quadratic baggage. As a result, compute scales linearly and peak memory remains constant. In experiments, we show that Delethink can be applied directly to off-the-shelf reasoning models ranging from $1.5\textnormal{B}$ to $30\textnormal{B}$ parameters, with no loss in performance. Extended reasoning becomes possible under fixed memory and linear compute, while enabling efficient RL training on new tasks. On the DeepScaleR dataset, Delethink trains R1DistillQwen1.5B to the same benchmark performance as a standard long chain-of-thought (LongCoT) approach, where both models generate up to $24\textnormal{k}$ thinking tokens. The difference is efficiency. Delethink reasons $40\%$ faster with $70\%$ less memory footprint. By decoupling reasoning length from context length, the Markovian Thinking paradigm opens the door to next-generation reasoning LLMs that can scale to millions of tokens with linear compute and constant memory.

Reinforcement learning (RL) has become a central paradigm for post-training large language models (LLMs) to elicit stronger reasoning. Yet, most recent RL for LLMs (RL4LLM) methods avoid explicit critics, replacing them with average advantage baselines. This shift is largely pragmatic: conventional value functions are computationally expensive to train at LLM scale and often fail under sparse rewards and long reasoning horizons. We revisit this bottleneck from an architectural perspective and introduce Asymmetric Proximal Policy Optimization (**AsyPPO**), a simple and scalable framework that restores the critic’s role while remaining efficient in large-model settings. **AsyPPO** employs a set of lightweight *mini-critics*, each trained on disjoint prompt shards. This design encourages diversity while preserving calibration, reducing value-estimation bias. Beyond robust estimation, **AsyPPO** leverages inter-critic uncertainty to refine the policy update: (i) masking advantages in states where critics agree and gradients add little learning signal, and (ii) filtering high-divergence states from entropy regularization, suppressing spurious exploration. Across multiple reasoning benchmarks, **AsyPPO** consistently improves learning stability and performance over strong baselines, e.g., GRPO, achieving performance gains of $> 6$% on *Qwen3-4b-Base* and about $3$% on *Qwen3-8b-Base* and *Qwen3-14b-Base* over classic PPO. Such results highlight the importance of architectural innovations in critics for scalable, efficient algorithms.

Scaling reinforcement learning (RL) to diverse multitask settings remains a central challenge. While recent advances in model-based RL achieve strong performance, they rely on planning and complex training pipelines, making it unclear which components are essential for scalability. We revisit this question and argue that the primary driver of scalable multitask RL is not model-based control, but representation learning. In particular, we show that combining predictive, model-based representations with high-capacity value function approximation is sufficient to achieve strong performance, even without planning. We evaluate a simple model-free algorithm, MR.Q, coupled with auxiliary predictive objectives into a scalable actor-critic architecture. This approach outperforms a recent world-model-based method and a range of deep RL baselines across a diverse suite of multitask continuous control tasks, while significantly reducing computational overhead and improving wall-clock efficiency. We observe consistent improvements with increased model capacity and show through ablations that predictive representation learning is critical for performance.

The quadratic cost of attention-based sequence models for long contexts has motivated a growing line of research on memory-based models that can compress context into a compact state. However, most existing memory models expose a static memory throughout the entire sequence. Because early tokens face no compression pressure, they occupy too many degrees of freedom and "pollute" the memory state, leaving little capacity for later context and increasing interference between what is stored and what arrives next. We study a new paradigm of incremental memory activation, where the effective capacity of memory is progressively expanded as the context grows. Imposing an early bottleneck forces the model to compress history more effectively, while unlocking fresh capacity over time reduces interference and improves retention of later context. We instantiate this paradigm in Proteus, a straightforward mechanism that can be incorporated into a broad class of neural memory architectures at no additional cost. We apply Proteus to state-of-the-art models, including SWLA, Comba, Titans, and Hope-Attention, and observe consistent improvements on standard language modeling and reasoning, as well as on long-context retrieval and understanding, with gains that grow at longer context lengths. Overall, our results show that static memory is suboptimal and that scheduling effective capacity is a simple and broadly applicable tool for sequence modeling.

In recent years, with the advancement of frontier AI, we have observed certain dynamics in open-sourcing and closed-sourcing decisions. We propose a game-theoretic model to analyze these dynamics in the current landscape of the AI race. Our model builds on an R D race framework under a winner-takes-all setting, and it accounts for the cases where the players' actions can be either discrete or continuous (i.e., partial open-sourcing, such as open weights). We show that determining the existence of a discrete pure non-trivial Nash equilibrium is NP-hard in general but that we can transform the discrete Nash existence computation into a MIP (Mixed-Integer Programming) problem, making it tractable for small instances using a standard MIP solver. Next, we show the existence and tractability of pure Nash equilibria in the continuous version of our problem, leveraging standard convex analysis results, and constructing an equivalent MIP formulation. Throughout this work, we leverage both our main technical results as well as surrounding technical analysis, to derive socially relevant insights that we believe can serve both to understand already existing decisions and dynamics and to potentially inform new policies.

Modern language models, including transformer, recurrent, and memory-based variants, share a common chassis: a stack of identical layers in which parameters are allocated uniformly across depth. This is a default inherited from the original transformer and largely unchanged since, yet a growing body of evidence suggests that layers contribute non-uniformly to the final output, with later layers refining the residual stream rather than transforming it. We ask whether parameter capacity should reflect this asymmetry. Our controlled experiment shows that, under a fixed budget, allocating more capacity to earlier layers and less to later layers improves perplexity over a uniform-width baseline, while the reverse allocation hurts. Building on this result, we introduce Tapered Language Models (TLMs), an architectural principle in which a parameter-bearing component is monotonically tapered across depth under a fixed total budget. MLPs are the natural site for this instantiation: they dominate parameter count across all modern LM families and expose width as a single, clean axis of variation. Across three model scales and four architectures (Transformer, Gated Attention, Hope-attention, and Titans), tapering MLP width via a smooth cosine schedule consistently improves perplexity and downstream benchmark performance over uniform baselines, at no additional parameter or compute cost. These findings establish depth-aware capacity allocation as a simple, architecture-agnostic axis of language model design, a free lever hidden in plain sight.

Reasoning has become a central capability in large language models. Recent research has shown that reasoning performance can be improved by looping an LLM's layers in the latent dimension, resulting in looped reasoning language models. Despite promising results, few works have investigated how their internal dynamics differ from those of standard feedforward models. In this paper, we conduct a mechanistic analysis of the latent states in looped language models, focusing in particular on how the stages of inference observed in feedforward models compare to those observed in looped ones. To this end, we analyze cyclic recurrence and show that for many of the studied models each layer in the cycle converges to a distinct fixed point; consequently, the recurrent block follows a consistent cyclic trajectory in the latent space. We provide evidence that as these fixed points are reached, attention-head behavior stabilizes, leading to constant behavior across recurrences. Empirically, we discover that recurrent blocks learn stages of inference that closely mirror those of feedforward models, repeating these stages in depth with each iteration. We study how recurrent block size, input injection, and normalization influence the emergence and stability of these cyclic fixed points. We believe these findings help translate mechanistic insights into practical guidance for architectural design.

Improving sample efficiency remains a core challenge in reinforcement learning (RL), especially in real-world settings like robotics, where data collection is costly. This challenge is pronounced in visual RL, where high-dimensional inputs often obscure learning signals. While prior work in visual RL has focused on algorithmic solutions, such as better dynamics models or exploration strategies, recent advances in state-based RL show that architectural design alone can lead to significant gains in sample efficiency. This raises an important question: Can these architectural principles transfer to visual RL? In response, we introduce V-Simba, a simple yet effective visual RL architecture inspired by the Simba architecture from state-based RL. Built on top of Soft Actor-Critic (SAC) with data augmentation, V-Simba modifies the architecture by adding normalization layers to stabilize training and using pointwise convolutions to reduce computation. Despite its simplicity, V-Simba matches or outperforms the state-of-the-art methods across the DMC, Adroit, and Meta-World benchmarks, while being more computationally efficient than DrQ-v2. We make our code publicly available at https://github.com/DAVIAN-Robotics/V-Simba.

Deep reinforcement learning systems often suffer from unstable training dynamics due to non-stationarity, where learning objectives and data distributions evolve over time. We show that under non-stationary targets, isotropic Gaussian embeddings are provably advantageous. In particular, they induce stable tracking of time-varying targets for linear readouts, achieve maximal entropy under a fixed variance budget, and encourage a balanced use of all representational dimensions. Building on this insight, we propose the use of Sketched Isotropic Gaussian Regularization for shaping representations toward an isotropic Gaussian during training. We demonstrate empirically, over a variety of domains, that this simple and computationally inexpensive method improves performance under non-stationarity while reducing representation collapse, neuron dormancy, and training instability.

Most applications of generative AI involve a sequential interaction in which a person inputs a prompt and waits for a response, and where reaction time and adaptivity are not important factors. In contrast, live jamming is a collaborative interaction that requires real-time coordination and adaptation without access to the other player's future moves, while preserving diversity to sustain a creative flow. Reinforcement learning post-training enables effective adaptation through on-policy interaction, yet it often reduces output diversity by exploiting coherence-based rewards. This collapse, known as “reward hacking”, affects many RL post-training pipelines, but is especially harmful in live jamming, where musical creativity relies on dynamic variation and mutual responsiveness. In this paper, we propose a novel adversarial training method on policy-generated trajectories to mitigate reward hacking in RL post-training for melody-to-chord accompaniment. A co-evolving discriminator separates policy trajectories from the data distribution, while the policy maximizes the discriminator output in addition to coherence rewards to prevent collapse to trivial outputs. We evaluate accompaniment quality and output diversity in simulation with both fixed test melodies and learned melody agents, and we conduct a user study with the model deployed in a real-time interactive system with expert musicians. Quantitative evaluation and user feedback demonstrate improved output diversity, harmonic coherence, adaptation speed and user agency. Our results demonstrate a simple yet effective method to mitigate reward hacking in RL post-training of generative sequence models.

While Proximal Policy Optimization (PPO) demonstrates strong performance in stationary settings, we show that its standard optimization paradigm struggles in continual and non-stationary environments. The failure does not stem from insufficient model capacity or overly restrictive clipping. Instead, PPO performs persistent, directionally inefficient local updates, which indicates a lack of geometry-aware guidance for accumulating meaningful behavioral change and ultimately hindering transitions toward new behavior patterns. Although divergence-based regularization introduces partial geometric awareness, its monotonically increasing penalties implicitly discourage large policy deviations, even when such shifts are necessary for effective adaptation. To address this limitation, we propose Gaussian Trust Region Policy Optimization (GTR), which reshapes the trust region using a Gaussian kernel. The resulting constraint is bounded and non-monotonic, providing strong local stability while progressively relaxing under sustained high-advantage updates. To further improve robustness, we introduce a Mixture Gaussian Anchor that adapts to recent policy trajectories, reducing variance induced by stale references. GTR is architecture-agnostic and achieves strong performance across games, simulated robotic control, open-world exploration, and language model post-training. These results demonstrate that geometry-aware trust-region design can be a promising direction for robust reinforcement learning in complex non-stationary environments. Our code is available at https://anonymous.4open.science/r/GTR_demo/README.md.

On-policy self-distillation achieves strong pass@1 accuracy by using a single model as both teacher and student, with the teacher conditioned on a correct demonstration to provide dense token-level feedback. We show that this could come at a hidden cost: rollout diversity decreases and pass@k curves flatten (i.e., generating more rollouts fails to improve accuracy). We trace this to compounding biases in the design of self-distillation with sampled demonstrations. The teacher scores each student rollout while conditioned on a sampled correct rollout, channeling its feedback through the model's own biases. We theoretically analyze the optimal self-distillation policy and show that it tilts the base distribution by a pointwise conditional mutual information score between the student's rollout and the correct rollout used as context. Unlike the ideal optimal on-policy reinforcement learning (RL), which preserves probability ratios among equally correct rollouts, self-distillation can amplify existing probability gaps, concentrating mass on already-dominant modes. On a controlled graph path-finding task and science question-answering benchmarks, self-distilled models match or exceed RL on average performance but exhibit substantially lower functional and semantic diversity, failing on out-of-distribution settings that require diverse strategies.

Recent works have proposed accelerating the wall-clock training time of actor-critic methods via the use of large-scale environment parallelization; unfortunately, these can sometimes still require large number of environment interactions to achieve a desired level of performance. Noting that well-structured representations can improve the generalization and sample efficiency of deep reinforcement learning (RL) agents, we propose the use of simplicial embeddings: lightweight representation layers that constrain embeddings to simplicial structures. This geometric inductive bias results in sparse and discrete features that stabilize critic bootstrapping and strengthen policy gradients. When applied to FastTD3, FastSAC, and PPO, simplicial embeddings consistently improve sample efficiency and final performance across a variety of continuous- and discrete-control environments, without any loss in runtime speed.

Artificially intelligent agents are increasingly being integrated into human decision-making: from large language model (LLM) assistants to autonomous vehicles. These systems often optimize their individual objective, leading to conflicts, particularly in general-sum games where naive reinforcement learning agents empirically converge to Pareto-suboptimal Nash equilibria. To address this issue, opponent shaping has emerged as a paradigm for finding socially beneficial equilibria in general-sum games. In this work, we introduce Advantage Alignment, a family of algorithms derived from first principles that perform opponent shaping efficiently and intuitively. We achieve this by aligning the advantages of interacting agents, increasing the probability of mutually beneficial actions when their interaction has been positive. We prove that existing opponent shaping methods implicitly perform Advantage Alignment. Compared to these methods, Advantage Alignment simplifies the mathematical formulation of opponent shaping, reduces the computational burden and extends to continuous action domains. We demonstrate the effectiveness of our algorithms across a range of social dilemmas, achieving state-of-the-art cooperation and robustness against exploitation.

Reinforcement learning (RL) has recently become a strong recipe for training reasoning LLMs that produce long chains of thought (LongCoT). Yet the standard RL "thinking environment", where the state is the prompt plus all prior reasoning tokens, makes the state unbounded and forces attention-based policies to pay quadratic compute as thoughts lengthen. We revisit the environment itself. We propose Markovian Thinking, a paradigm in which the policy advances reasoning while conditioning on a constant-size state, decoupling thinking length from context size. As an immediate consequence this yields linear compute with constant memory. We instantiate this idea with Delethink, an RL environment that structures reasoning into fixed-size chunks. Within each chunk, the model thinks as usual; at the boundary, the environment resets the context and reinitializes the prompt with a short carryover. Through RL, the policy learns to write a textual state near the end of each chunk sufficient for seamless continuation of reasoning after reset. Trained in this environment, an R1-Distill 1.5B model reasons in 8K-token chunks yet thinks up to 24K tokens, matching or surpassing LongCoT-RL trained with a 24K budget. With test-time scaling, Delethink continues to improve where LongCoT plateaus. The effect of linear compute is substantial: we empirically estimate at 96K average thinking length LongCoT-RL costs 27 H100-months vs. 7 for Delethink. Analysis at RL initialization shows off-the-shelf reasoning models (1.5B-120B) often sample Markovian traces zero-shot across diverse benchmarks, providing positive samples that make RL effective at scale. Our results show that redesigning the thinking environment is a powerful lever: it enables very long reasoning without quadratic overhead and opens a path toward efficient, scalable reasoning LLMs.

The widespread adoption of generative AI models has raised growing concerns about representational harm and potential discriminatory outcomes. Yet, despite growing literature on this topic, the mechanisms by which bias emerges - especially in unconditional generation - remain disentangled. We define the bias of an attribute as the difference between the probability of its presence in the observed distribution and its expected proportion in an ideal reference distribution. In our analysis, we train a set of unconditional image generative models and adopt a commonly used bias evaluation framework to study bias shift between training and generated distributions. Our experiments reveal that the detected attribute shifts are small. We find that the attribute shifts are sensitive to the attribute classifier used to label generated images in the evaluation framework, particularly when its decision boundaries fall in high-density regions. Our empirical analysis indicates that this classifier sensitivity is often observed in attributes values that lie on a spectrum, as opposed to exhibiting a binary nature. This highlights the need for more representative labeling practices, understanding the shortcomings through greater scrutiny of evaluation frameworks, and recognizing the socially complex nature of attributes when evaluating bias.