AI papers — 2026-10-19

The Morning Preprint — this day as one edition: the day's papers, the five things worth knowing, and the episodes behind them.

Today we are diving into how we can make large language models much more memory efficient when fine-tuning them fully. This is crucial because fitting these massive models onto available hardware is a real bottleneck for training.

The work that matters most here is TACO, which proposes a Ternary Absolute-max Column-wise One-sparse optimizer. This optimizer drastically cuts down on optimizer state memory, reducing it by about 174 times compared to AdamW8bit while maintaining comparable accuracy and runtime on models like OPT-13B.

This efficiency gain is significant because it allows us to fine-tune much larger models. It enables full-parameter fine-tuning of 30 to 32 billion parameter models on a single eighty gigabyte H100 GPU. This optimization method builds upon the geometric ideas explored in Muon, which also reduces optimizer memory through matrix-valued updates. However, TACO's specific approach is designed to avoid performance degradation when starting from AdamW-pretrained models.

Moving down the list of importance, we have CBPO. This method unifies learning from offline demonstrations and on- and off-policy rollouts by making them compatible through a shared regression loss. This allows optimization to retain the same objective without needing separate SFT or RL objectives, achieving leading performance across various reasoning benchmarks.

This is complemented by State-OPD, which grounds teacher supervision in interaction states for multi-turn agentic environments. This approach shows that it achieves the highest macro-average success rate among methods for models like Qwen3-4B.

Finally, we have work focused on improving robustness and structure. Robust Langevin Sampling under State-Dependent Heavy-Tailed Noise with Applications to LLM Fine-Tuning establishes convergence guarantees for sampling methods even when gradients are not globally Lipschitz. This is vital for Bayesian fine-tuning of LLMs.

Additionally, Monarch-MoE introduces structured experts using block-diagonal factors in LoRA to combat catastrophic forgetting during continual instruction tuning by isolating gradient updates between blocks.

The work by Haiyuan Wan Yanyi Li Weicheng Lin Cong Fang on Sample-Adaptive Zeroth-Order Post-Training is particularly relevant. It tackles how to efficiently fine-tune large language models without relying on uniform sampling, which is a major bottleneck in current methods. This paper proposes Sample-Adaptive Zeroth-Order Optimization, or SAZO.

SAZO learns sample importance online by extracting gradient norm information from directional differences to guide queries adaptively. This means the method doesn't just look at every example equally; it focuses its computational effort where it matters most for faster convergence.

The core mechanism involves maintaining per-example exponential moving averages to track how the importance of different samples evolves as training progresses. This approach is significant because SAZO proves that it matches the leading convergence term of the minimum-variance sampling oracle while achieving strictly faster convergence rates in terms of query count across several model families.

STAPLE, introduced by Ranajoy Sadhukhan et al., addresses the scalability issue in sparse computation for pretraining. It introduces Static Sparse Upcycling of Pretrained Checkpoints via Embedding Modules. This method integrates static sparse memory into pretrained models, specifically targeting Mixture-of-Experts models where expert homogeneity is a problem.

STAPLE improves performance on benchmarks like HumanEval and MBPP by learning complementary token-specific information while preserving the original pretrained computation. It also achieves faster decoding speeds with less memory usage.

This idea of leveraging static capacity to augment pretrained knowledge connects somewhat to PHBA, which introduces Prefix-State Hybrid Block Attention. PHBA aims to balance long-context modeling with precise token retrieval.

PHBA replaces local sliding window attention with top-k block-sparse retrieval coupled with a compact prefix state summarizing preceding context. This allows the model to combine long-range evidence with historical context within a unified layer.

On a different front, the clinical application of AI detection is highlighted by Expert-level detection of low-voltage intervals in EEG as a robust biomarker of neonatal encephalopathy. The ConvNeXt model developed for this task achieved an AUC of 0.996 and expert-level agreement. This suggests that AI can reliably identify specific physiological markers that quantitative features alone cannot capture.

Finally, the theoretical work on online optimization provides context for how learning can proceed sequentially in uncertain environments, such as Optimal and Efficient Online Inverse Optimization. This algorithm establishes a deterministic approach for learning an unknown linear objective with a regret bound of O(sqrt d). This shows that polynomial-time solutions exist even when the objective function is never directly observed.

The most significant work presented today involves developing a framework for robust importance sampling to estimate rare event probabilities. This is crucial because standard methods often fail in high-dimensional or multimodal settings where they become unstable. This approach separates the problem into coverage and fitting stages to overcome initial barriers. It constrains the final proposal distribution to ensure finite variance, meaning the estimation can be practically stable even with limited sampling budgets.

This work builds upon earlier efforts by proposing a framework that substantially improves efficiency over various baselines like crude Monte Carlo or variational inference methods. This is complemented by research into selective test-time adaptation for multimodal time series, where CARAT proposes decoupling model reliance from runtime corruption detection to guide how missing data is handled during deployment.

Another important contribution comes from the development of a transferable graph-learning framework called G-R2R. It reframes road network disruption as a region-conditioned forecasting problem. This allows it to be applied to unseen transportation networks without extensive retraining, contrasting with methods that rely on local calibration for flood impact models.

Furthermore, there is work on optimizing nuclear criticality experiments using physics-informed neural networks to design new geometries that maximize similarity scores with target technologies. This optimization technique allows for searching over complex material combinations within a grid space.

The most significant work this morning is Trajectory-Aware Policy Optimization by Cai et al. This addresses the problem of state distribution shift in long-horizon reinforcement learning for large language models. When an LLM generates long sequences, local policy mismatches can lead to severe shifts in the state distribution over time, causing training instability and collapse.

TAPO proposes a principled framework that derives this correction from the off-policy policy gradient and incorporates it into the RL objective. It uses a geometric-mean estimator for variance reduction and length-aware dynamic clipping to prevent overly deviated updates across trajectories.

This work builds upon earlier ideas in LLM reinforcement learning by focusing on trajectory dynamics rather than just token or sequence level mismatches. It achieves substantial gains, showing up to forty-eight percent improvement over GRPO on mathematical reasoning and algorithmic coding benchmarks while significantly enhancing training stability.

This focus on trajectory awareness is complemented by CB-Orchestrator, which tackles the problem of manually designing complex LLM agent workflows. CB-Orchestrator decouples workflow generation from selection, using a Contextual Bandit problem to sample efficiently and adaptively select optimal workflows during inference. This reduces training overhead by over forty percent on the MATH dataset.

Another area where structure matters is in how we handle representation space for neural networks. Neighborhood Smoothing for Calibration introduces graph smoothing as a general principle to encourage similar predictive distributions across neighboring samples in representation space. This helps combat overconfidence and miscalibration in modern neural networks.

This method proposes a graph-based regularizer that penalizes the Jensen--Shannon divergence between the predictive distributions of neighboring samples, showing it improves predictive quality and complements post-hoc calibration techniques.

In terms of model expressivity, research into Random-Feature MPNNs shows that continuous random node features can be quantized to finitely many colors with arbitrarily small uniform error in expected output for specific architectures. This reduction allows researchers to transfer limitations on finite-color distinguishability to limitations on size-uniform approximation when using these continuous random features.

This contrasts with the analysis done by AKMPC, which examines how continuous random features affect the expressivity of message-passing neural networks and establishes incomparability between random-color MPNNs and higher-order Weisfeiler-Leman methods.

The topic of routing for multimodal models is also advancing, as MHR-Router proposes a framework for Vision-Language Models to adaptively dispatch queries to the most suitable expert using local retrieval evidence. This system fuses frozen multimodal features into a query representation and then retrieves expert success and failure memories as local evidence to select the optimal model. This shows consistent accuracy improvements over traditional routing baselines.

Finally, in forecasting applications, Diffusion Precipitation Generation with Medium-range Weather Forecasting tackles the bias present in existing weather foundation models trained on reanalysis data. It uses a separate generator to map non-precipitation forecasts to observation-aligned daily precipitation. This framework uses a multi-scale recurrent encoder and cross-attention during denoising, leading to improved daily precipitation accuracy and event detection compared to global ensemble evaluations.

The work on opening the black box of classifier-free guidance matters because it provides a principled way to control when conditional information is most useful during diffusion model generation. This moves beyond arbitrary schedules, which current methods rely on heuristics for. Understanding the underlying dynamics allows for better inference-time control.

The paper AKMPC explores this by framing the diffusion trajectory as an information bottleneck and finding that guidance effectiveness peaks after a specific Fisher-information balance condition is met. This insight motivates the proposal of a training-free dynamic CFG strategy that activates guidance based on estimated guidance gain crossing a threshold. This aligns injection with generation dynamics, recovering interval-based guidance as a special case.

Moving down in significance, the work FedPCD addresses fairness issues in federated learning by separating global optimization from client protection. It allows only minimum necessary corrections for client constraints while preserving the primary objective. This framework consistently outperforms baselines on heterogeneous benchmarks and shows faster convergence by reaching performance targets in fewer communication rounds.

EchoTab introduces a measurement-grounded representation learning framework for echocardiography that explicitly models measurements as an intermediate step between views and diagnosis. It uses measurement-specific tokens to selectively attend to relevant views, which improves classification accuracy by aligning learned attention with clinical measurement relationships.

MS-ECG-FM develops a universal electrocardiogram foundation model by training it through contrastive alignment against multiple distinct clinical note types, including reports from echocardiography and radiology. This multi-source alignment captures complementary information across different diagnostic domains, leading to stronger representations for various cardiac conditions than models trained on single sources.

Finally, DecepEval introduces a benchmark for evaluating deception in LLM agents by framing it within the LLM Deception Diamond framework based on pressure, incentive, opportunity, and conflict. This allows researchers to systematically measure how external conditions induce deception across different models and task families.

The most important work here is RUBRIC because it tackles a real problem in risk-sensitive areas like fraud detection where minority class samples are rare. Existing oversampling methods often create low-quality synthetic data that just makes things worse.

This framework introduces a way to select synthetic samples based on a quality-over-quantity optimization. It uses a neural density ratio discriminator to judge realism against real minority samples and a margin-based scoring function g-(t)=- (1+e-t/τ) to measure how close the sample is to the decision boundary.

This filtering approach helps tighten generalization bounds for margin-based classifiers by simultaneously reducing distribution shift and suppressing contributions from near-negative tails. This is significant because it improves recall while keeping overall discriminative ability intact across different data generators.

This idea of balancing quality and quantity in synthetic generation relates to how set-valued policy learning handles ambiguity in recommendations. The latter aims to output sets of valuable treatments rather than a single arbitrary intervention.

Set-Valued Policy Learning explores situations where multiple treatments might be equally good or have no effect, proposing that outputting sets of valuable treatments better supports clinical decision-making by reflecting the recommendation's ambiguity. This concept connects to how models learn from complex data, similar to how the COntextual Retrieval and Encoding model attempts to build a single cognitive core from disparate principles of cognition.

The work on CacheScope is also relevant because it focuses on acceleration in diffusion models by intelligently deciding when to use full computation versus cached approximations. This mirrors the need for careful selection in sample generation methods like RUBRIC.

Furthermore, DiffusionART introduces a method for efficiently training diffusion models using aggregated rewards, stabilizing training without needing complex EMA tuning.

Finally, the findings from NV-Tesseract-AD suggest that structuring reconstruction difficulty along two axes can improve anomaly detection performance in time series data by guiding the learning process through structured corruption. This structured approach to learning relates to how provenance-grouped protocols are used in perovskite photovoltaics to separate memorization from true physical generalization.

The most critical work today is SepsisLens because it tackles the fundamental challenge of predicting deterioration in intensive care units while maintaining the necessary structural integrity for clinical interpretation. It addresses how to detect dangerous changes from noisy patient measurements without losing the connection between an alert and the specific physiological signals that caused it.

This approach is vital because many existing temporal models collapse these variable-specific connections into a single patient-level risk score, which hinders clinicians' ability to understand what is actually driving the warning.

SepsisLens achieves this by preserving variable-indexed temporal states until the point of risk composition. This allows for observation-aware representations that encode each variable's dynamics separately while sharing a temporal encoder for trajectories AKMPC. This structure enables the StructuredRiskHead to compose multi-horizon risk from explicit variable and organ-level components, which is a significant improvement over models that only provide scalar risk prediction.

The geometric rediscovery of singular value decomposition by AKMPC offers a deeper mathematical foundation for many machine learning techniques. It shows that the machinery behind principal component analysis and kernel methods is rooted in maximizing stretch across dimensions. This concept underpins how we can analyze the feature importance seen in SepsisLens, as input-side masking showed that ranked components reflected variables with greater influence on prediction.

GeneICL provides a powerful tabular foundation model for bulk transcriptomics, showing that transcriptomics-aware pretraining can outperform purely self-supervised models by incorporating measured expression profiles into a recurrent architecture AKMPC. This work suggests that the quality of the underlying data structure matters more than just scaling up the model size for clinical prediction tasks.

BeatFlow-ECG demonstrates a successful method for reconstructing single-channel ECG from indirect wearable signals by using conditional rectified flow. It achieves high correlation metrics when incorporating motion information AKMPC. This shows how complex signal reconstruction can be handled effectively with explicit conditioning on external factors like movement.

Learning Cross-Model Activation Alignments with Explicit Many-to-Many Layer Maps explores how to understand the relationship between independently trained large language models by learning maps that align their internal activations AKMPC. This is important for understanding feature transfer between different AI systems.

Targeted search in wave-based neural operators suggests that searching for worst-case error devices can provide a lower bound on how large errors can become under fabrication variation AKMPC. This provides a necessary caution when deploying hardware solutions for physical simulations.

The Tree SAE work is most significant because it directly tackles the fundamental challenge of properly learning hierarchical structures in complex data like large language models. It moves beyond simple activation coverage which can lead to false connections between unrelated concepts.

This method introduces a reconstruction constraint alongside activation constraints to enforce a deeper functional link between feature levels. This approach significantly outperforms existing Sparse Autoencoders when learning hierarchical pairs while still matching state-of-the-art performance on several benchmarks. It provides a way to map the geometry of child feature subspaces to uncover complex concept structures within these models.

This is more powerful than the prior work by Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders, which only used activation coverage as a condition for hierarchy.

The SOAP framework is also important because it allows for failure attribution in LLM agents without needing costly step-level labels or success trajectories. It effectively identifies the decisive steps causing task failure through spectral scoring and attention propagation. This contrasts with methods that rely on external annotations.

Regarding policy updates, the analysis of Ascent Conditions for Length-Normalized Single-Rollout Policy Updates shows that dividing trajectory scores by their realized length can change the direction of a reward-centered policy update. This demonstrates how this normalization affects learning in finite trajectory models. This is related to the findings in Ascent Conditions for Length-Normalized Single-Rollout Policy Updates, which found that length normalization can cause an expected Euclidean update to decrease task success under certain conditions.

The study on Rethinking Relative Policy Differences Transfer of On-Policy Distillation clarifies that the type of policy difference used for supervision matters. Differences arising from task-specific training transfer consistently, whereas those based purely on model scale transfer weakly or not at all. This is a nuanced finding compared to the general idea behind Rethinking Relative Policy Differences Transfer of On-Policy Distillation, which developed a unified view for using proxy experts and references in on-policy distillation.

Finally, the research into Privileged Context as Drift in On-Policy Self-Distillation suggests that how privileged context is chosen—whether its content or source is varied—affects the stability and direction of policy drift. This indicates that context design should be part of stability considerations for continual learning. This relates to the broader theme explored in Privileged Context as Drift in On-Policy Self-Distillation.

Today's papers

The papers

Important terms

TACO
Ternary Absolute-max Column-wise One-sparse optimizer is a new method that drastically reduces optimizer state memory, cutting it down by about 174 times compared to AdamW8bit while keeping accuracy similar.
Sample-Adaptive Zeroth-Order Optimization (SAZO)
SAZO learns sample importance online by looking at gradient norm information from directional differences. This guides the model to focus its training effort on the most important examples for faster convergence.
Trajectory-Aware Policy Optimization
This framework corrects state distribution shifts in long-horizon reinforcement learning by incorporating corrections derived from off-policy policy gradients, leading to better stability and performance.
SepsisLens
This model predicts deterioration in intensive care units by preserving variable-indexed temporal states. It keeps track of each patient's specific physiological signals, which is vital for clinical understanding.