Measuring the in-context computational effort of language models is a key challenge, as metrics like next-token loss fail to capture reasoning complexity. Prior methods based on latent state compressibility can be invasive and unstable. We propose Multiple Token Divergence (MTD), a simple measure of computational effort defined as the KL divergence between a model's full output distribution and that of a shallow, auxiliary prediction head. MTD can be computed directly from pre-trained models with multiple prediction heads, requiring no additional training. Building on this, we introduce Divergence Steering, a novel decoding method to control the computational character of generated text. We empirically show that MTD is more effective than prior methods at distinguishing complex tasks from simple ones. On mathematical reasoning benchmarks, MTD correlates positively with problem difficulty. Lower MTD is associated with more accurate reasoning. MTD provides a practical, lightweight tool for analyzing and steering the computational dynamics of language models.

Unified multimodal models (UMMs) aim to jointly perform multimodal understanding and generation within a single framework. We present TUNA, a native UMM that builds a unified continuous visual representation by cascading a VAE encoder with a representation encoder. This unified representation space allows end-to-end processing of images and videos for both understanding and generation tasks. Compared to prior UMMs with decoupled representations, TUNA's unified visual space avoids representation format mismatches introduced by separate encoders, outperforming decoupled alternatives in both understanding and generation. Moreover, we observe that stronger pretrained representation encoders consistently yield better performance across all multimodal tasks, highlighting the importance of the representation encoder. Finally, in this unified setting, jointly training on both understanding and generation data allows the two tasks to benefit from each other rather than interfere. Our extensive experiments on multimodal understanding and generation benchmarks show that TUNA achieves state-of-the-art results in image and video understanding, image and video generation, and image editing, demonstrating the effectiveness and scalability of its unified representation design.

We propose a new frontier: Neural Computers (NCs) – an emerging machine form that unifies computation, memory, and I/O in a learned runtime state. Unlike conventional computers, which execute explicit programs, agents, which act over external execution environments, and world models, which learn environment dynamics, NCs aim to make the model itself the running computer. Our long-term goal is the Completely Neural Computer (CNC): the mature, general-purpose realization of this emerging machine form, with stable execution, explicit reprogramming, and durable capability reuse. As an initial step, we study whether early NC primitives can be learned solely from collected I/O traces, without instrumented program state. Concretely, we instantiate NCs as video models that roll out screen frames from instructions, pixels, and user actions (when available) in CLI and GUI settings. These implementations show that learned runtimes can acquire early interface primitives, especially I/O alignment and short-horizon control, while routine reuse, controlled updates, and symbolic stability remain open. We outline a roadmap toward CNCs around these challenges. If overcome, CNCs could establish a new computing paradigm beyond today's agents, world models, and conventional computers.

We introduce MoS (Mixture of States), a novel fusion paradigm for multimodal diffusion models that merges modalities using flexible, state-based interactions. The core of MoS is a learnable, token-wise router that creates denoising timestep- and input-dependent interactions between modalities' hidden states, precisely aligning token-level features with the diffusion trajectory. This router sparsely selects the top-k hidden states and is trained with an ε-greedy strategy, efficiently selecting contextual features with minimal learnable parameters and negligible computational overhead. We validate our design with text-to-image generation (MoS-Image) and editing (MoS-Editing), which achieve state-of-the-art results. With only 3B to 5B parameters, our models match or surpass counterparts up to 4× larger. These findings establish MoS as a flexible and compute-efficient paradigm for scaling multimodal diffusion models.

The attention mechanism in a Transformer architecture matches key to query based on both content—the what—and position in a sequence—the where. We present an analysis indicating that what and where are entangled in the popular rotary position embedding (RoPE). This entanglement can impair performance particularly when decisions require independent matches on these two factors. We propose an improvement to RoPE, which we call Polar Coordinate Position Embedding or PoPE, that eliminates the what-where confound. PoPE is far superior on a diagnostic task requiring indexing solely by position or by content. On autoregressive sequence modeling in music, genomic, and natural language domains, Transformers using PoPE as the positional encoding scheme outperform baselines using RoPE with respect to evaluation loss (perplexity) and downstream task performance. On language modeling, these gains persist across model scale, from 124M to 774M parameters. Crucially, PoPE shows strong zero-shot length extrapolation capabilities compared not only to RoPE but even a method designed for extrapolation, YaRN, which requires additional fine tuning and frequency interpolation.

Large language models (LLMs) face a fundamental trade-off between computational efficiency (e.g., number of parameters) and output quality, especially when deployed on computationally limited devices such as phones or laptops. One way to address this challenge is by following the example of humans and have models ask for help when they believe they are incapable of solving a problem on their own; we can overcome this trade-off by allowing smaller models to respond to queries when they believe they can provide good responses, and deferring to larger models when they do not believe they can. To this end, in this paper, we investigate the viability of Predict-Answer/Act (PA) and Reason-Predict-Reason-Answer/Act (RPRA) paradigms where models predict – prior to responding – how an LLM judge would score their output. We evaluate three approaches: zero-shot prediction, prediction using an in-context report card, and supervised fine-tuning. Our results show that larger models (particularly reasoning models) perform well when predicting generic LLM judges zero-shot, while smaller models can reliably predict such judges well after being fine-tuned or provided with an in-context report card. Altogether, both approaches can substantially improve the prediction accuracy of smaller models, with report cards and fine-tuning achieving mean improvements of up to 55

Editing the figures in a research paper is a routine and time-consuming part of everyday research practice: authors relabel components, rearrange panels, and restyle visuals as they revise their manuscripts. Automating this editing workflow under a natural-language instruction, however, is challenging, because a scientific figure is a dense infographic in which heterogeneous visual elements such as schematics, plots, photos, captions, and arrows are composed under a tight visual grammar to advance a specific argument. To address this, we present SciDiagramEdit, a benchmark and skill-evolution framework that learns from natural paper revisions and operates on the figure's editable vector source, where users can inspect and co-edit individual primitives alongside the agent. Our benchmark mines before/after figure pairs from arXiv version histories, each grounded in the authors' own revision intent. To accommodate the diversity of editing instructions, we adopt agentic learning via skill evolution: an agentic proposer continually refines the agent's skill specification from execution traces over multiple epochs. The resulting skill progressively lifts edit accuracy on a held-out validation set, providing evidence that natural paper revisions are an effective training signal for instruction-driven figure editing.

Self-improving autonomous agents are moving from research prototypes to deployed systems. The primary goal is controllable evolution, or adaptation, from experience with minimal or even no human input. This survey frames modern self-improving agents as adaptive systems that convert experience into accumulated capability gains. We offer a system-level framework that represents a modern agent as a configuration coupling a foundation model with an operational scaffold of prompts, memory, tools, and control logic. Within this framework, self-improvement is formalized as a self-induced update operator that obtains and commits updates to model parameters or scaffold components. We organize prior work by update target and by the signals that drive change, then review applications and discuss evaluation, before closing with open problems and future directions. For convenience, we track technical updates on https://github.com/selfimproving-agent/awesome-Self-Improving-Agents.

Many robotic platforms expose an API through which external software can command their actuators and read their sensors. However, transitioning from these low-level interfaces to high-level autonomous behaviour requires a complicated pipeline, whose components demand distinct areas of expertise. Existing approaches to bridging this gap either require retraining for every new embodiment or have only been validated across structurally similar platforms. We introduce RACAS (Robot-Agnostic Control via Agentic Systems), a cooperative agentic architecture in which three LLM/VLM-based modules (Monitors, a Controller, and a Memory Curator) communicate exclusively through natural language to provide closed-loop robot control. RACAS requires only a natural language description of the robot, a definition of available actions, and a task specification; no source code, model weights, or reward functions need to be modified to move between platforms. We evaluate RACAS on several tasks using a wheeled ground robot, a recently published novel multi-jointed robotic limb, and an underwater vehicle. RACAS consistently solved all assigned tasks across these radically different platforms, demonstrating the potential of agentic AI to substantially reduce the barrier to prototyping robotic solutions.

Recent studies operationalize self-improvement through coding agents that edit their own codebases, grow a tree of self-modifications through expansion strategies that favor higher software engineering benchmark performance, considering that this implies more promising subsequent self-modifications. However, we identify a mismatch between the agent’s self-improvement potential (metaproductivity) and its coding benchmark performance, namely the \emph{Metaproductivity-Performance~Mismatch}. Inspired by Huxley’s concept of clade, we propose a metric ($\mathrm{CMP}$) that aggregates the benchmark performances of the \emph{descendants} of an agent as an indicator of its potential for self-improvement. We show that the G\"odel Machine, the optimal self-improving machine, is achieved with access to true $\mathrm{CMP}$. We introduce the Huxley-G\"odel Machine (HGM), which, by estimating $\mathrm{CMP}$ and using it as guidance, searches the tree of self-modifications. On SWE-bench Verified and Polyglot, HGM outperforms prior self-improving coding agent search methods while using less wall-clock time. Moreover, the agent optimized by HGM on SWE-bench Verified outperforms SWE-agent, a leading human-engineered open source coding agent on SWE-bench Lite, where SWE-agent ranks the best on the official leaderboard, when both use the GPT-5-mini backbone, demonstrating that HGM self-improvement indeed enhances genuine coding capability.

AI-based computer-aided synthesis planning (CASP) systems are in demand as components of AI-driven drug discovery workflows. However, the high latency of such CASP systems limits their utility for high-throughput synthesizability screening in de novo drug design. We propose a transformer-based single-step retrosynthesis model with reduced inference latency based on speculative beam search combined with a scalable drafting strategy called Medusa. Replacing the standard transformer and beam search with our approach accelerates the expansion stage of the planning algorithm, leading to higher solvability in CASP when planning under stringent time limits, and saves hours of computation when synthesis is constrained by the number of iterations. Our method brings AI-based CASP systems closer to meeting the stringent latency requirements of high-throughput synthesizability screening and improving the overall user experience.

Large Vision-Language Models (VLMs) now act as agents in interactive environments, where success requires coherent reasoning and decision-making across turns. Although end-to-end training in agentic environments can improve such multi-turn decision-making abilities, current methods mainly rely on either token-wise optimization over concatenated token trajectories or turn-wise optimization with uniform within-turn credit. In this work, we establish theoretical formulations for the two levels of optimization and derive a hybrid advantage that serves both objectives. Furthermore, with an appropriate choice of discount factor and learning target, we prove that a unified critic model can estimate values for both turn-wise and token-wise. As such, we propose HyGAE, an actor-critic framework that jointly optimizes token- and turn-level objectives with the hybrid advantage and unified critic. We conduct extensive evaluations of HyGAE across five multi-turn decision-making environments, where it achieves an average success rate of 91% and a significant improvement of 10% over other methods. Furthermore, we provide an in-depth analysis showing that the exact analytic form of the hybrid advantage and return is crucial for optimization. Project Page: https://wx-zhang.github.io/hygae-web/.

Morphology-control co-design concerns the coupled optimization of an agent's body structure and control policy. This problem exhibits a bi-level structure, where the control dynamically adapts to the morphology to maximize performance. Existing methods typically neglect the control's adaptation dynamics by adopting a single-level formulation that treats the control policy as fixed when optimizing morphology. This can lead to inefficient optimization, as morphology updates may be misaligned with control adaptation. In this paper, we revisit the co-design problem from a game-theoretic perspective, modeling the intrinsic coupling between morphology and control as a novel variant of a Stackelberg game. We propose Stackelberg Proximal Policy Optimization (Stackelberg PPO), which explicitly incorporates the control's adaptation dynamics into morphology optimization. By modeling this intrinsic coupling, our method aligns morphology updates with control adaptation, thereby stabilizing training and improving learning efficiency. Experiments across diverse co-design tasks demonstrate that Stackelberg PPO outperforms standard PPO in both stability and final performance, opening the way for dramatically more efficient robotics designs.

One of the bottlenecks on the way towards recursively self-improving systems is the challenge of interestingness: the ability to prospectively identify which tasks or data hold the potential for future progress. We formalize interestingness as an inductive heuristic for future compression progress and investigate its predictability using tools from Kolmogorov Complexity and Algorithmic Statistics. By analyzing complexity-runtime profiles under Length, Algorithmic, and Speed priors, we demonstrate that the inductive property of interestingness – the capacity for past progress to signal future discovery – is theoretically viable and empirically supported. We prove that expected future progress depends exponentially on the recency of the last observed breakthrough. Furthermore, we show that the Algorithmic Prior is significantly more optimistic than the Length Prior, yielding a quadratic increase in expected discovery for the same observed profile. These findings are experimentally confirmed across three diverse universal computational paradigms.

This position paper argues that truly open-ended intelligence is bottlenecked by the challenge of : the ability to prospectively identify which tasks or data hold the potential for future progress. We formalize interestingness as an inductive heuristic for future compression progress and investigate its predictability using tools from Kolmogorov Complexity and Algorithmic Statistics. By analyzing complexity-runtime profiles under various priors over computable objects, we demonstrate that the —the capacity for past compression progress to signal future discovery—is theoretically viable. However, we show that this property is highly sensitive to the underlying distribution of objects. We conclude by calling for a move beyond human-in-the-loop filtering or data creation, and a shift toward models that can explicitly assess their own potential for insight. Furthermore, we advocate the engineering of scale-free synthetic environments, providing a principled roadmap for the development of truly autonomous open-ended systems.

The use of LLMs for code generation has naturally extended to code testing and evaluation. As codebases grow in size and complexity, so does the need for automated test generation. Current approaches for LLM-based test generation rely on strategies that maximize immediate coverage gain, a greedy approach that plateaus on code where reaching deep branches requires setup steps that individually yield zero new coverage. Drawing on principles of Bayesian exploration, we treat the program's branch structure as an unknown environment, and an evolving coverage map as a proxy probabilistic posterior representing what the LLM has discovered so far. Our method, CovQValue, feeds the coverage map back to the LLM, generates diverse candidate plans in parallel, and selects the most informative plan by LLM-estimated Q-values, seeking actions that balance immediate branch discovery with future reachability. Our method outperforms greedy selection on TestGenEval Lite, achieving 51-77

Continual learning agents with finite capacity must balance acquiring new knowledge with retaining the old. This requires controlled forgetting of knowledge that is no longer needed, freeing up capacity to learn. Weight decay, viewed as a mechanism for forgetting, can serve this role by gradually discarding information stored in the weights. However, a fixed scalar weight decay drives this forgetting uniformly over time and uniformly across all parameters, even when some encode stable knowledge while others track rapidly changing targets. We introduce Forgetting through Adaptive Decay (FADE), which adapts per-parameter weight decay rates online via approximate meta-gradient descent. We derive FADE for the online linear setting and apply it to the final layer of neural networks. Our empirical analysis shows that FADE automatically discovers distinct decay rates for different parameters, complements step-size adaptation, and consistently improves over fixed weight decay across online tracking and streaming classification problems.

Discovery—formulating novel generalizations—is a central part of the scientific process. Despite its importance, there is a gap in the current AI benchmark landscape, with few benchmarks directly probing the capacity for discovering new knowledge with experimentation in controlled environments where the objective is unknown. To address this gap, we release a new benchmark: DiG-bench (Discovery in Games). DiG-bench consists of a set of 70 independent games. Each game is encoded as a short string and has unique transformation rules that must be discovered through interaction and experimentation. The levels of the game present a series of challenges to test whether the rules have been discovered, where the win conditions for each level are also unknown. We provide games at seven tiers of difficulty for AI agents. The lowest tier is routinely solvable by multiple models, while the highest tier challenges the best models in agentic harnesses. All 70 games were solved by at least one human on first attempt. A subset of 21 games is released publicly, and the remainder is held private for secure evaluation.

Upside Down Reinforcement Learning (UDRL) is a promising framework for solving reinforcement learning problems which focuses on learning command-conditioned policies. In this work, we extend UDRL to the task of learning a command-conditioned generator of deep neural network policies. We accomplish this using Hypernetworks - a variant of Fast Weight Programmers, which learn to decode input commands representing a desired expected return into command-specific weight matrices. Our method, dubbed Upside Down Reinforcement Learning with Policy Generators (UDRLPG), streamlines comparable techniques by removing the need for an evaluator or critic to update the weights of the generator. To counteract the increased variance in last returns caused by not having an evaluator, we decouple the sampling probability of the buffer from the absolute number of policies in it, which, together with a simple weighting strategy, improves the empirical convergence of the algorithm. Compared with existing algorithms, UDRLPG achieves competitive performance and high returns, sometimes outperforming more complex architectures. Our experiments show that a trained generator can generalize to create policies that achieve unseen returns zero-shot. The proposed method appears to be effective in mitigating some of the challenges associated with learning highly multimodal functions. Altogether, we believe that UDRLPG represents a promising step forward in achieving greater empirical sample efficiency in RL. A full implementation of UDRLPG is publicly available at https://github.com/JacopoD/udrlpg_

The act of telling stories is a fundamental part of what it means to be human. This work introduces the concept of narrative information, which we define as the overlap in information space between a story and the items that compose the story. Using contrastive learning methods, we show how modern artificial neural networks can be leveraged to distill stories and extract a representation of the narrative information. We then demonstrate how evolutionary algorithms can leverage this to extract a set of narrative template curves and how these-in tandem with a novel curve-fitting algorithm we introduce-can reorder music albums to automatically induce stories in them. In doing so, we give statistically significant evidence that (1) these narrative information template curves are present in existing albums and that (2) people prefer an album ordered through one of these learned template curves over a random one. The premises of our work extend to any form of (largely) independent media, and as evidence, we also show that our method works with image data.