AI papers — 2026-10-18

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

The work by Harish Haresamudram Wanting Mao Aditya Radhakrishnan Maxwell A. Xu James M. Rehg introduces VarMAE, a variable input masked autoencoder designed to learn one representation that works across different sampling rates and temporal contexts by creating two views of the same signal and using a cross-resolution masking objective AKMPC. This approach is significant because it shows that a single, parameter-efficient foundation model can achieve broad transfer capabilities across different sensing configurations, even when pretraining data has fixed input constraints.

Another important area involves improving how we learn from complex network structures by understanding how small changes propagate through them. VQIL develops a trainable quantum walk for seed-conditioned ranking on graphs VQIL, which uses quantum interference to rank nodes based on their response strength to a perturbation. This shows significant improvements over classical models in both power-flow and protein-protein interaction network scenarios.

This contrasts with the work by Taeyoung Yu Seonbin Jo Jiwon Kim Junyoung Byun who created SeoulMMOD, a large-scale benchmark for origin-destination flow forecasting. This benchmark addresses the limitations of existing benchmarks by providing multimodal, multi-year data for citywide mobility predictions.

Finally, Mason Minot Simon Friedensohn Cédric R. Weber propose dynamic budget allocators for generative condition selection in de novo nanobody design Inference-Time Steering Across Scaffolds and Epitopes for Sample-Efficient De Novo Nanobody Design. This work empirically showed nearly five times the yield of designable candidates per GPU-hour by steering compute across different configurations.

This efficiency gain is complemented by Sparse-TD-JEPA Sparse-TD-JEPA: Improving Zero-Shot Reinforcement Learning With Sparse Latent Task Representations. This method introduces an l1 sparsity penalty to improve zero-shot reinforcement learning performance by concentrating reward structure into a smaller latent subspace.

The study by Arthur da Cunha Kasper Green Larsen Liang-Yu Zou AKMPC shows that boosting weak hypotheses up to strong ones depends heavily on the structure of the base class, which they quantify using a gamma-VC dimension. This parameter measures how complex a learning problem is in terms of sample size needed for weak-to-strong learning. They sharpen the link between this new gamma-VC dimension and the classic VC dimension.

This finding is significant because it gives us a more precise tool to understand when boosting will actually work well, moving beyond just general bounds. Relatedly, Richard Cornelius Suwandi Feng Yin Kevin Murphy’s Kernel Autoresearch Kernaut tackles the problem of designing good kernels for machine learning models by treating kernel design as open-ended model discovery.

They found that agents can discover reusable inductive biases in these kernels, and a human-refined version of one could improve predictive error by 5.7%.

In terms of representation learning, Van Dai Do Huu Hiep Nguyen Minh Hoang Nguyen Manh Nguyen Hung Le Deakin’s Lotus LOTUS introduces a pretraining principle called forecast grounding for time-series language models. This requires the model to predict the future evolution of a signal before adapting it to downstream tasks. They showed this method, using a temporal-spectral patch representation, achieves high accuracy with very few trainable parameters.

Meanwhile, exploring how models adapt continuously is addressed by Rishabh Tiwari Kusha Sareen Lakshya A Agrawal’s Learning Fast and Slow FST framework. This proposes separating model parameters into "slow" weights that retain general knowledge and "fast" context weights that learn task-specific information. This is shown to be more sample-efficient than standard reinforcement learning while reducing catastrophic forgetting.

The work on Gated Adaptive Positional Encoding is particularly important because it directly addresses the instability that arises when large language models handle sequences longer than what they were trained on. This approach introduces a content-aware bias into the attention mechanism to selectively suppress irrelevant distant context while protecting salient tokens, allowing these models to generalize much further than before.

This concept builds upon existing work that only partially tackles these failures by trading local resolution for stability. The GAPE framework achieves this by using a query-dependent gate to contract unnecessary context and a key-dependent gate to keep important distant tokens accessible. This mechanism is integrated directly into standard scaled dot-product attention, meaning it can be retrofitted into existing models like the 7B model mentioned in the GAPE description.

In terms of memory construction for agents, Preping offers a way to build procedural memory before any task experience is available by using a proposer-guided framework. This involves a Proposer generating synthetic tasks conditioned on a structured control state, which is then executed by a Solver and validated by a Validator to selectively insert trajectories into memory. The main advantage comes from the proposer's control over feasibility and coverage, leading to deployment costs that are significantly lower than online memory construction methods.

This idea of structured control over generated data synthesis is related to learning how to pose problems for training reasoning models. The problem generator in Learning to Pose Problems explicitly reasons about problem directions before synthesizing data. It adapts difficulty based on the solver's ability and uses a reward signal based on answer consistency to guide this process. This method has shown performance gains over seed-only training across various reasoning benchmarks.

The most critical work here is EconCausal because it directly tackles how large language models can reason about cause and effect when different contexts change. This paper introduces EconCausal, a massive benchmark built from 10,490 context-annotated causal triplets drawn from over 2,595 high-quality economic studies. The core issue is whether an LLM can correctly infer the direction of a causal effect given a specific context and then adjust that inference when the context shifts.

The results show that while top models achieve 84% accuracy in predicting effects within a single context, their mean accuracy drops significantly to 74.75% overall. Accuracy plummets to just 44.27% when the signs of the source and target inputs disagree. Furthermore, accuracy falls below 60% when the source signs are corrupted, indicating sensitivity to input quality.

Another piece of work that touches on how models handle reasoning paths is CoTS. This framework reframes latent reasoning not as a binary choice but as a continuous allocation between discrete commitment and soft exploration at each step. It uses the competition between the two most likely candidates to dynamically balance these modes, which consistently outperforms standard Chain-of-Thought methods across six different reasoning benchmarks.

Moving to optimization methods, PISO offers a way to tackle problems where you only have partial gradient information by combining it with a running estimate of the residual error. This technique allows for randomized sketch iterations with explicit contraction mechanisms. It provides convergence guarantees that depend on the residual's variation rather than needing uniform bounds on the unavailable gradient component.

We also see work exploring adaptive decision-making in uncertain environments through AdaSwitch, which is a meta-algorithm designed to switch between offline and online oracle methods for history-dependent problems. This framework provides explicit performance guarantees that tighten as prediction errors decrease or the offline optimum increases, showing applicability in areas like lead-time quotation.

Finally, UniFGVC addresses the problem of overfitting when fine-tuning multimodal large language models on few-shot fine-grained visual classification tasks by shifting to a training-free retrieval paradigm. It uses an attribute-aware module to generate structured text descriptions that pinpoint discriminative visual attributes. This allows the model to recognize classes by comparing features against category templates rather than relying on sparse samples.

The most significant development today concerns OrBIT, a framework that learns reusable local geometry from orbit dynamics to guide embedding compression in large language models. This matters because it shows how to drastically reduce the storage footprint of these massive models while maintaining high performance. OrBIT achieves remarkable results, reaching 37.9 times compression on GPT-2 and over twenty-three times compression on each 7B table relative to 16-bit storage, while still matching established quantization techniques in rate distortion performance.

The mechanism behind this is quite elegant: the global reconstruction residual decides where the fixed coding budget is spent. Redundant overlapping charts allow local errors to compensate after gluing together. This structure-guided compression contrasts with other work like Distilling Graph Geometry: Knowledge Gap from GNNs to MLPs, which focuses on retaining predictive accuracy when moving from graph neural networks to graph-free MLPs by using Ollivier-Ricci curvature to guide supervision.

Another area of interest is how models handle temporal fidelity, as seen in Moment-Video, which introduced a benchmark to diagnose whether video multimodal large language models can reliably use brief, answer-critical visual evidence. This work highlights that even when sampling covers all necessary event intervals, substantial errors persist on longer videos, suggesting temporal localization remains a challenge for these systems.

On the agent side, there is progress in runtime adaptation with Life-Harness Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents. This improves frozen LLM agents by evolving training trajectories into reusable interventions across environment contracts. This method shows that adapting the harness rather than just updating model parameters can lead to substantial accuracy gains, such as exceeding Meta-Harness and AHE by significant margins with models like Qwen3-4B-Instruct.

Furthermore, optimizing agent skills is being tackled by MOCHA MOCHA: Multi-Objective Chebyshev Annealing for Agent Skill Optimization. This uses Chebyshev scalarization to select skills that improve the Pareto front of accuracy and economy simultaneously. This technique has shown that agents can escape their initial seed skill in nearly every run, achieving gains up to twenty-three percent on specific benchmarks by discovering trade-offs an average would never favor.

The most pressing issue is how we can actually measure whether large language models are useful in complex professional settings, which is where BankerToolBench comes in by creating a benchmark for end-to-end analytical workflows in investment banking. This matters because if we cannot reliably evaluate agent performance against real job tasks, the progress of AI adoption remains purely theoretical.

BankerToolBench requires agents to perform tasks like navigating data rooms and generating multi-file deliverables such as Excel models and pitch decks, which takes bankers up to twenty-one hours to complete. Testing nine frontier models showed that even GPT-5.4 failed nearly half of the rubric criteria, with bankers rating zero percent of its outputs as client-ready, pointing toward major problems like breakdowns in cross-artifact consistency.

This difficulty in evaluation is compounded by the general challenge of predicting model performance before experiments through anticipatory evaluation, where EVALCAST provides a resource for studying how much can be inferred from natural language descriptions of tasks. This relates to the practical deployment concerns addressed by AEGIS, which creates an operational infrastructure for governing adaptive medical AI under regulations like the Predetermined Change Control Plan.

AEGIS introduces a framework with three modules—dataset assimilation and retraining, model monitoring, and conditional decision—to operationalize regulatory concepts like those from the FDA and EU AI Act. For instance, when testing sepsis prediction models across eleven simulated iterations, AEGIS correctly identified eight APPROVE decisions but also flagged critical states where no deployable model existed while the released one was simultaneously at risk.

Furthermore, ensuring safety in these deployments involves certified robustness. Certified Robustness via Sparse Second-Order Cone Relaxations introduces a hierarchy of second-order cone programming relaxations to provide stronger certificates for neural networks than simpler linear programming methods. This method showed that class-optimized robust evaluation can match the accuracy of SDP certification while significantly reducing verification time.

These technical safety measures are also relevant when considering how models handle complex inputs. Masked Attention Pooling Is Invariant to Padding and Permutation proves that pooling instance embeddings is robust to padding and permutation under ideal real arithmetic. This contrasts with the practical issues seen in BankerToolBench where consistency breaks down, linking the need for rigorous mathematical guarantees to the messy reality of agent execution.

Today's papers

The papers

Important terms

VarMAE
A variable input masked autoencoder that learns a single representation usable across different sampling rates and temporal contexts by creating two views of a signal and using a cross-resolution masking objective.
VQIL
Uses quantum interference to rank nodes on graphs based on how strongly they respond to small perturbations, showing improvements over classical models in areas like power flow and protein interaction networks.
EconCausal
A benchmark testing if large language models can correctly infer the direction of cause and effect when context changes, highlighting their sensitivity to input quality and context shifts.
OrBIT
A framework that learns reusable local geometry from orbit dynamics to guide embedding compression in large language models, drastically reducing storage needs while maintaining high performance.