Iterative

Persistent AI memory is often reduced to a retrieval problem: store prior interactions as text, embed them, and ask the model to recover relevant context later. This design is useful for thematic recall, but it is mismatched to the kinds of memory that agents need in production: exact facts, current state, updates and deletions, aggregation, relations, negative queries, and explicit unknowns. These operations require memory to behave less like search and more like a system of record. This paper argues that reliable external AI memory must be schema-grounded. Schemas define what must be remembered, what may be ignored, and which values must never be inferred. We present an iterative, schema-aware write path that decomposes memory ingestion into object detection, field detection, and field-value extraction, with validation gates, local retries, and stateful prompt control. The result shifts interpretation from the read path to the write path: reads become constrained queries over verified records rather than repeated inference over retrieved prose. We evaluate this design on structured extraction and end-to-end memory benchmarks. On the extraction benchmark, the judge-in-the-loop configuration reaches 90.42% object-level accuracy and 62.67% output accuracy, above all tested frontier structured-output baselines. On our end-to-end memory benchmark, xmemory reaches 97.10% F1, compared with 80.16%-87.24% across the third-party baselines. On the application-level task, xmemory reaches 95.2% accuracy, outperforming specialised memory systems, code-generated Markdown harnesses, and customer-facing frontier-model application harnesses. The results show that, for memory workloads requiring stable facts and stateful computation, architecture matters more than retrieval scale or model strength alone.

Recent work on large language models (LLMs) has emphasized the importance of scaling inference compute. From this perspective, the state-of-the-art method Scattered Forest Search (SFS) has been proposed, employing Monte Carlo Tree Search with carefully crafted initial seeds and textual optimization for multi-turn program correction. However, its complexity makes it unclear what factors contribute to improvements in inference performance. To address this problem, we analyze SFS and propose a simpler method, \textsc{Iterative Refinement of Repair Instructions} (IRRI), which fixes initial programs and iteratively refines repair instructions. Because of the simplicity of IRRI, we theoretically establish the non-elimination of IRRI using Oracle-Guided Inductive Synthesis (OGIS). Experiments on several program generation benchmarks suggest that IRRI achieves inference performance comparable to state-of-the-art methods. These results indicate that, even without complex search structures, refining initial programs with high-quality repair instructions alone can effectively improve inference performance.

Retrieval-Augmented Generation (RAG) grounds language models in external evidence, but multi-hop question answering remains difficult because iterative pipelines must control what to retrieve next and when the available evidence is adequate. In practice, systems may answer from incomplete evidence chains, or they may accumulate redundant or distractor-heavy text that interferes with later retrieval and reasoning. We propose S2G-RAG (Structured Sufficiency and Gap-judging RAG), an iterative framework with an explicit controller, S2G-Judge. At each turn, S2G-Judge predicts whether the current evidence memory supports answering and, if not, outputs structured gap items that describe the missing information. These gap items are then mapped into the next retrieval query, producing stable multi-turn retrieval trajectories. To reduce noise accumulation, S2G-RAG maintains a sentence-level Evidence Context by extracting a compact set of relevant sentences from retrieved documents. Experiments on TriviaQA, HotpotQA, and 2WikiMultiHopQA show that S2G-RAG improves multi-hop QA performance and robustness under multi-turn retrieval. Furthermore, S2G-RAG can be integrated into existing RAG pipelines as a lightweight component, without modifying the search engine or retraining the generator.

Batch processes are inherently transient and typically nonlinear, motivating nonlinear model predictive control (NMPC). However, adopting NMPC is hindered by the cost and unavailability of dynamic models. Thus, we propose to use Gaussian Processes (GP) in a model-learning NMPC scheme (GP-MLMPC) for batch processes. We initialize the GP-MLMPC using data from a single initial trajectory, e.g., from a PI controller. We iteratively apply the NMPC embedded with GPs to run batches and update the GP with new observations from each iteration, thereby achieving batch-wise improvements. Using uncertainty quantification from the GPs, we formulate chance constraints to enforce safe operation to the required confidence levels. We demonstrate our approach in \textit{silico} on a semi-batch polymerization reactor for tracking and economic objectives over durations of two hours, and the reactor temperature is constrained in a range of $\pm2^\circ C$ around its setpoint. After only four batch iterations, tracking error from the GP-MLMPC scheme converged to a reduction of $83\%$, compared to the initial trajectory. Furthermore, under an economic objective, the GP-MLMPC resulted in a 17-fold increase in final product mass by iteration 8, compared to the initial trajectory. In both cases, the resulting GP-MLMPC performance is on par with the full-model NMPC, which shows that the optimal controller can be learned by the approach. By collecting samples around the optimal trajectory, the GP-MLMPC remains sample-efficient across iterations and achieves quick convergence. Thus, the proposed GP-MLMPC scheme presents a promising data-efficient approach for the control of nonlinear batch processes without mechanistic knowledge.

Normalizing Flows (NFs) are a classical family of likelihood-based methods that have received revived attention. Recent efforts such as TARFlow have shown that NFs are capable of achieving promising performance on image modeling tasks, making them viable alternatives to other methods such as diffusion models. In this work, we further advance the state of Normalizing Flow generative models by introducing iterative TARFlow (iTARFlow). Unlike diffusion models, iTARFlow maintains a fully end-to-end, likelihood-based objective during training. During sampling, it performs autoregressive generation followed by an iterative denoising procedure inspired by diffusion-style methods. Through extensive experiments, we show that iTARFlow achieves competitive performance across ImageNet resolutions of 64, 128, and 256 pixels, demonstrating its potential as a strong generative model and advancing the frontier of Normalizing Flows. In addition, we analyze the characteristic artifacts produced by iTARFlow, offering insights that may shed light on future improvements. Code is available at https://github.com/apple/ml-itarflow.

Trajectory Optimization (TO) solvers exploit known system dynamics to compute locally optimal trajectories through iterative improvements. A downside is that each new problem instance is solved independently; therefore, convergence speed and quality of the solution found depend on the initial trajectory proposed. To improve efficiency, a natural approach is to warm-start TO with initial guesses produced by a learned policy trained on trajectories previously generated by the solver. Diffusion-based policies have recently emerged as expressive imitation learning models, making them promising candidates for this role. Yet, a counterintuitive challenge comes from the local optimality of TO demonstrations: when a policy is rolled out, small non-optimal deviations may push it into situations not represented in the training data, triggering compounding errors over long horizons. In this work, we focus on learning-based warm-starting for gradient-based TO solvers that also provide feedback gains. Exploiting this specificity, we derive a first-order loss for Sobolev learning of diffusion-based policies using both trajectories and feedback gains. Through comprehensive experiments, we demonstrate that the resulting policy avoids compounding errors, and so can learn from very few trajectories to provide initial guesses reducing solving time by $2\times$ to $20 \times$. Incorporating first-order information enables predictions with fewer diffusion steps, reducing inference latency.

Tabular language models can generate synthetic tables by modeling rows as token sequences, but they are typically trained once with supervised fine-tuning and then used as static synthesizers. This is limiting because next-token likelihood does not directly optimize the distributional, utility, and indistinguishability properties used to evaluate synthetic data. We study iterative reward-guided post-training for tabular language models through a generate--score--align protocol, where a generator samples synthetic rows, a task-specified reward ranks them, and the model is updated relative to a fixed supervised reference. Within this protocol, we propose \textbf{TabGRAA} (\textbf{Tab}ular \textbf{G}roup-\textbf{R}elative \textbf{A}dvantage \textbf{A}lignment), a group-relative alignment method that compares high- and low-reward generated groups using group-averaged policy/reference log-ratios rather than one-to-one preference pairs. Across five mixed-type benchmarks, TabGRAA improves a GReaT backbone beyond additional supervised fine-tuning and achieves the strongest average trade-off among adapted DPO, KTO, and NPO baselines on fidelity and downstream utility, while maintaining empirical privacy diagnostics near the supervised baseline. Ablations show that the gains depend on meaningful reward ranking and stable group-level updates rather than extra training alone. Reward-substitution and scorer-separation studies further show that the post-training loop can use both classifier-based and classifier-free rewards, and that proper scorer separation is important for preserving the fidelity--utility--privacy trade-off. These results position TabGRAA as a self-improving post-training method for tabular language-model generators, complementary to strong static tabular synthesizers.

Looped transformers scale computational depth without increasing parameter count by repeatedly applying a shared transformer block and can be used for iterative refinement, where each loop rewrites a full fixed-size prediction in parallel. On difficult problems, such as those that require search-like computation, reaching a highly structured solution starting from noise can require long refinement trajectories. Learning such trajectories is challenging when training specifies only the target solution and provides no supervision over the intermediate refinement path. Diffusion models tackle this issue by corrupting data with varying magnitudes of noise and training the model to reverse it in a \textit{single step}. However, this process misaligns training and testing behaviour. We introduce Denoising Recursion Models, a method that similarly corrupts data with noise but trains the model to reverse the corruption over \textit{multiple} recursive steps. This strategy provides a tractable curriculum of intermediate states, while better aligning training with testing and incentivizing non-greedy, forward-looking generation. Through extensive experiments, we show this approach outperforms the Tiny Recursion Model (TRM) on ARC-AGI, where it recently achieved breakthrough performance.

Recursive architectures such as Tiny Recursive Models (TRMs) perform implicit reasoning through iterative latent computation, yet the geometric structure of these reasoning trajectories remains poorly understood. We investigate the activation manifold of TRMs during recursive unrolling and find that activations occupy an effectively linear, low-dimensional subspace whose principal directions can be tracked dynamically with cheap power iterations. This suggests that weight-sharing concentrates iterative computation along a small number of dominant eigendirections, and we find that this concentration varies sharply across computational sites. We exploit this structure through LASER (Low-Rank Activation SVD for Efficient Recursion), a dynamic compression framework that maintains an evolving low-rank basis via matrix-free subspace tracking with a fidelity-triggered reset mechanism, achieving ${\sim}60\%$ activation memory savings with no statistically significant accuracy degradation. Our analysis raises questions about how recursive architectures allocate representational capacity during implicit reasoning, and whether this concentration can be exploited to improve the efficiency and stability of latent computation.

Efficient and stable solution of partial differential equations (PDEs) is central to scientific and engineering applications, yet existing numerical solvers rely heavily on matrix based discretizations, while learning based methods require costly training and often suffer from limited generalization. In this work, we proposes a PDE energy driven framework that solves PDEs through physically constrained diffusion iterations, without relying on classical matrix based finite element assembly or data driven neural network training. The proposed method evolves arbitrary random initial fields through PDE energy driven implicit iterations combined with Gaussian smoothing, while strictly enforcing boundary conditions at each iteration. The proposed formulation is applied to representative one dimensional Poisson, Heat, and viscous Burgers equations, covering both steady state and transient problems. Numerical results demonstrate stable convergence to the unique physical solution from random initializations, with accurate resolution of sharp gradients and controlled Mean Squared Error (MSE) across a wide range of discretization parameters. Detailed comparisons with analytical solutions indicate that the framework achieves competitive accuracy and stability. Overall, the proposed framework provides a fast, flexible, and physically consistent alternative to traditional numerical solvers, offering a potential pathway for scalable PDE solutions in both research and engineering applications.

In many real-world settings, problem instances that need to be solved are quite similar, and knowledge from previous optimization runs can potentially be utilized. We explore this for the Traveling Salesperson problem with time windows (TSPTW), which often arises in settings where the travel-time matrix is fixed but time-window constraints change across related tasks. Existing TSPTW studies, however, have not systematically compared solving such task sequences independently with sequential transfer from previously solved tasks. We address this gap using a multi-task benchmark in which each base instance is expanded into five related tasks under two environments: partial time-window expansion and swap-additive time reassignment. We compare a standard from-scratch protocol with an iterative protocol that initializes each task from the best tour of the previous task, using the popular local search approaches LNS, VNS, and LKH-3 under a common penalized-score objective. Our experimental results show that the iterative protocol is consistently superior in the progressive-relaxation setting and generally competitive under swap-additive changes, with improvements increasing on more difficult instances.

Reinforcement learning with verifiable rewards (RLVR) and Rubrics as Rewards (RaR) have driven strong gains in domains with clear correctness signals and even in subjective domains by synthesizing evaluation criteria from ideal reference answers. But many real-world tasks admit multiple valid outputs and lack the single ideal answer that rubric generation depends on. We identify this reference-free setting as a gap in current post-training methods and propose Implicit Error Counting (IEC) to fill it. Instead of checking what a response gets right against a rubric, IEC enumerates what it gets wrong, applying severity-weighted scores across task-relevant axes and converting them into calibrated per-aspect rewards. We show that naïve explicit enumeration is too noisy for stable optimization, and that two design choices: implicit score emission and group calibration are necessary to make error counting a reliable reward. As a case study, we validate IEC on virtual try-on (VTO), a domain that is simultaneously too constrained for holistic scoring and too permissive for rubric-based evaluation: subtle garment errors are unacceptable, yet many output variations are correct. We introduce Cascaded Error Counting (CEC) as an evaluation metric, which tracks human preferences well (60% top-1 vs. 30% others), and curate Mismatch-DressCode (MDressBench), a benchmark with maximal attribute mismatch to stress-test reward designs. On MDressBench, IEC outperforms RaR across all metrics (CEC: 5.31 vs. 5.60 on flat references; 5.20 vs. 5.53 on non-flat). On VITON-HD and DressCode, IEC matches or surpasses six baselines on 6 of 8 perceptual metrics. These results suggest that when ideal answers are unavailable, counting errors provide a stronger signal than constructing rubrics.

Generative sequence modeling faces a fundamental tension between the expressivity of Transformers and the efficiency of linear sequence models. Existing efficient architectures are theoretically bounded by shallow, single-step linear updates, while powerful iterative methods like Test-Time Training (TTT) break hardware parallelism due to two dimensions of serial dependency: token-level state reliance and step-level iteration loops. We propose PRISM (Parallel Residual Iterative Sequence Model) to resolve this tension. PRISM explicitly approximates the expressive gate-residual-direction iteration pattern of TTT in a parallelizable form. We employ a Write-Forget Decoupling strategy that isolates non-linearity within the injection operator. To bypass the serial dependency of explicit solvers, PRISM utilizes a two-stage proxy architecture: a short-convolution anchors the initial residual using local history energy, while a learned predictor estimates the refinement updates directly from the input. This design distills structural patterns associated with iterative correction into a parallelizable feedforward operator. Theoretically, we prove that this formulation achieves Rank-$L$ accumulation, structurally expanding the update scheme beyond the single-step Rank-$1$ bottleneck. Empirically, it achieves comparable performance to explicit optimization methods while achieving \textbf{174x higher throughput}. Codes are available in https://github.com/gpr-prism/prism/.

Learned image reconstruction has become a pillar in computational imaging and inverse problems. Among the most successful approaches are learned iterative networks, which are formulated by unrolling classical iterative optimisation algorithms for solving variational problems. While the underlying algorithm is usually formulated in the functional analytic setting, learned approaches are often viewed as purely discrete. In this survey we present a unified operator view for learned iterative networks. Specifically, we formulate a learned reconstruction operator, defining how to compute, and separately the learning problem, which defines what to compute. In this setting we present common approaches and show that many approaches are closely related in their core. We review linear as well as non-linear inverse problems in this framework and present a short numerical study to conclude.

Shadows encode rich information about scene geometry and illumination, yet existing methods either predict a unified shadow mask or overlook attached shadows entirely. We address this gap by proposing a framework for jointly detecting cast and attached shadows through explicit physical modeling of light direction and surface geometry under a dominant directional-light setting. Our approach is grounded in a simple observation: surfaces facing away from the light source tend to fall into shadow. We exploit the reciprocal relationship between shadow formation and light estimation to construct a closed feedback loop, a dual-module architecture in which a shadow detection module and a light estimation module iteratively refine each other. At each pass, updated light estimates, together with surface normals, produce partial attached shadow maps that guide detection, while improved shadow predictions sharpen light estimation. To support training and evaluation, we introduce a dataset of 1,458 images with manually annotated cast and attached shadow masks sourced from three existing benchmarks. Experiments demonstrate that our proposed method outperforms prior methods, with at least a 33% reduction in attached-shadow BER, while maintaining strong full-shadow and cast-shadow performance.

Practical LLM agents often operate over multi-turn conversations where success is determined only after the full interaction ends. Most multi-turn RL methods train via on-policy rollouts, but unlike in single-turn RLHF, the policy cannot produce a trajectory alone, since an external environment must respond after each agent turn. For conversational agents, this environment is a user, but real users are generally unavailable inside the training loop and simulated users are difficult to build faithfully. Separately, real-world deployment is rarely fully online or fully offline. The common production pattern is called "batch online," where the current policy is deployed to collect a batch of interaction data, then retrained on that batch and redeployed. We show that this batch-online setting creates an opportunity for applying the classical approximate policy iteration algorithm. Our central observation is that running standard token-level, single-turn GRPO with a learned turn-level $Q$-function as the reward model is a policy improvement step for the multi-turn problem. Building on this, we present Iterative GRPO, which alternates between fitting $Q^π$ from logged Monte Carlo returns via standard reward modeling (policy evaluation) and running single-turn GRPO against $Q^π$ (policy improvement). During policy improvement, the learned $Q^π$ scores candidate responses by their expected downstream return, avoiding the need for simulator rollouts of the remaining conversation. Overall, this allows us to do principled multi-turn RL using only single-turn RLHF methods without an interactive environment or user simulator inside the training loop. We demonstrate the effectiveness of Iterative GRPO on six multi-turn negotiation environments.

Multi-source stationary computed tomography (CT) has recently attracted attention for its ability to achieve rapid image reconstruction, making it suitable for time-sensitive clinical and industrial applications. However, practical systems are often constrained by ultra-sparse-view sampling, which significantly degrades reconstruction quality. Traditional methods struggle under ultra-sparse-view settings, where interpolation becomes inaccurate and the resulting reconstructions are unsatisfactory. To address this challenge, this study proposes Diffusion-Refined Neural Attenuation Fields (Diff-NAF), an iterative framework tailored for multi-source stationary CT under ultra-sparse-view conditions. Diff-NAF combines a Neural Attenuation Field representation with a dual-branch conditional diffusion model. The process begins by training an initial NAF using ultra-sparse-view projections. New projections are then generated through an Angle-Prior Guided Projection Synthesis strategy that exploits inter view priors, and are subsequently refined by a Diffusion-driven Reuse Projection Refinement Module. The refined projections are incorporated as pseudo-labels into the training set for the next iteration. Through iterative refinement, Diff-NAF progressively enhances projection completeness and reconstruction fidelity under ultra-sparse-view conditions, ultimately yielding high-quality CT reconstructions. Experimental results on multiple simulated 3D CT volumes and real projection data demonstrate that Diff-NAF achieves the best performance under ultra-sparse-view conditions.

Pruning is a core technique for compressing neural networks to improve computational efficiency. This process is typically approached in two ways: one-shot pruning, which involves a single pass of training and pruning, and iterative pruning, where pruning is performed over multiple cycles for potentially finer network refinement. Although iterative pruning has historically seen broader adoption, this preference is often assumed rather than rigorously tested. Our study presents one of the first systematic and comprehensive comparisons of these methods, providing rigorous definitions, benchmarking both across structured and unstructured settings, and applying different pruning criteria and modalities. We find that each method has specific advantages: one-shot pruning proves more effective at lower pruning ratios, while iterative pruning performs better at higher ratios. Building on these findings, we advocate for patience-based pruning and introduce a hybrid approach that can outperform traditional methods in certain scenarios, providing valuable insights for practitioners selecting a pruning strategy tailored to their goals and constraints. Source code is available at https://github.com/janumiko/pruning-benchmark.

Gibbs measures, such as Coulomb gases, are popular in modelling systems of interacting particles. Recently, we proposed to use Gibbs measures as randomized numerical integration algorithms with respect to a target measure $π$ on $\mathbb R^d$, following the heuristics that repulsiveness between particles should help reduce integration errors. A major issue in this approach is to tune the interaction kernel and confining potential of the Gibbs measure, so that the equilibrium measure of the system is the target distribution $π$. Doing so usually requires another Monte Carlo approximation of the \emph{potential}, i.e. the integral of the interaction kernel with respect to $π$. Using the methodology of large deviations from Garcia--Zelada (2019), we show that a random approximation of the potential preserves the fast large deviation principle that guarantees the proposed integration algorithm to outperform independent or Markov quadratures. For non-singular interaction kernels, we make minimal assumptions on this random approximation, which can be the result of a computationally cheap Monte Carlo preprocessing. For the Coulomb interaction kernel, we need the approximation to be based on another Gibbs measure, and we prove in passing a control on the uniform convergence of the approximation of the potential.

We study distributionally robust optimization (DRO) for robust inference when the worst-case distribution is continuous, leading to significant computational challenges due to the infinite-dimensional nature of the optimization problem. Unlike traditional discrete DRO approaches, which often suffer from scalability issues, limited generalization, and costly worst-case inference, our framework exploits Brenier's theorem to characterize the least favorable distribution as the pushforward of a transport map from a continuous reference measure. This characterization motivates our study of the minimax problem in Wasserstein space. We propose an iterative algorithmic framework with multiple variants and establish global convergence guarantees under mild assumptions, deriving complexity bounds in terms of subgradient evaluations and inexact Jordan-Kinderlehrer-Otto updates. Numerical results with neural network-based transport maps demonstrate that the proposed method enables both stable training of robust classifiers and effective worst-case inference for classification tasks.