AI papers — 2026-10-23

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

Today's work centers on understanding how large vision-language models can fail when they encounter inputs they don't understand or when the information they have is contradictory, which is crucial because these misbehaviors pose real risks in critical applications. The most important piece of research here is Evidential Uncertainty Quantification, or EUQ, which fine-tunes uncertainty measurement to capture both internal knowledge conflict and ignorance in a single model pass AKMPC. This method interprets the model's output features as evidence, modeling how conflicting or missing information leads to specific failures like hallucinations or out-of-distribution problems.

This idea of measuring logical consistency is also important, as standard accuracy metrics often miss when a model respects underlying rules, which is why we look at the Rule Violation Score Delplanque. This metric tells us if a model follows predefined constraints, showing that two models can be equally accurate but behave very differently logically.

We also see work on aligning these models to human preferences, such as Projected Alignment Reward Estimated from Demonstrations Inverse RL, which tries to derive explicit rewards from expert examples to guide post-training alignment. Furthermore, we are exploring how to create better evaluation sets for these agents; for instance, NodeSynth generates synthetic data designed with sociotechnical nuance to stress-test models in sensitive areas like medical advice NodeSynth. Finally, there is research into making tool calling reliable across modalities, showing how text-to-voice conversion affects performance for agents using tools From Text to Voice.

Medical concept extraction from electronic health records is crucial because it underpins many downstream applications, but existing benchmarks mainly focus on concepts that are explicitly stated rather than those that are implied in medical narratives. MedicalBench addresses this by formulating concept extraction as a verification task over note-concept pairs, paired with sentence-level evidence identification to assess both correctness and interpretability. This benchmark is built using MIMIC-IV discharge summaries and human-verified ICD one zero codes, curated through a multi stage large language model triage pipeline before expert review.

The performance of state of the art large language models on MedicalBench remains modest, which suggests that extracting implicitly expressed concepts is inherently difficult. Furthermore, the results indicated that performance is largely invariant to note length, implying that the difficulty lies in reasoning rather than superficial factors like text size. This difficulty in understanding implicit medical meaning connects to work on self derived rewards for relation extraction, which shows how using reinforcement learning intermediate rewards can improve accuracy and explanation quality by encouraging reasoning to align with labels.

In terms of agentic systems, the TabAgent framework offers a way to make complex decision making cheaper by replacing slow large language model calls with a compact textual tabular classifier trained on execution traces. This approach extracts structured schema and state features from trajectories, augments coverage with synthetic supervision, and then scores candidates using a lightweight head. TabAgent successfully maintained task level success on the long horizon AppWorld benchmark while cutting latency by about ninety five percent and inference cost by eighty five to ninety one percent.

Another area of focus is in image detection, where Align Don't Blend shows that using disentangled orthogonal supervision instead of blended supervision is key to generalization. This involves pairing a vision language model with purely semantic supervision and a vision expert with purely artifact supervision. Even with this simpler training recipe, the simplest detector achieved a nine point five percent improvement across five in-the-wild benchmarks when trained on only two generators.

Structural supervision is also important for medical foundation models, as seen in Multiview Masked Autoencoders which leverage the natural multi view organization of radiology studies to learn view invariant representations. Extending this with MVMAE V two t incorporates radiology reports as an auxiliary text based signal to boost semantic grounding, especially in low label regimes.

Finally, when looking at how models handle data generation, research into in context bias propagation demonstrates that even mild biases in in context examples can cause global statistical distortions when generating synthetic tabular data. This highlights a persistent challenge regarding the sensitivity of large language models to adversarial prompts used for bias injection.

The Mixture-of-Control approach is particularly significant because it tackles the problem of limited information flow in state-based fine-tuning methods. This technique treats block-wise control states as experts in a sparse mixture of experts process, which allows for efficient communication across transformer blocks, thereby enhancing representation learning during adaptation. This contrasts with prior methods where cross-block communication often incurred too much computational cost to be practical for efficient fine-tuning.

This efficiency is further supported by the work on WindINR, which demonstrates that updating only a low-dimensional latent state yields a fourteen-fold speedup over full-network fine-tuning while preserving continuous queries at arbitrary spatial coordinates. This shows how focusing updates on a specific latent state can drastically reduce computational load while maintaining functional capability.

Another important development is the proposal of Heavy Tailed Activation Function (HTAF), which provides a smooth approximation to the Heaviside function to enable stable training with gradient-based optimization. This stability is crucial because it allows for the training of models like Spiking Neural Networks and Deep Heaviside neural networks, while Implicit Concept Bottleneck Models (ICBMs) use this activation function to induce interpretable discrete feature representations.

RelAgent offers a different kind of adaptation by using Large Language Model agents as data scientists for relational learning. This system first uses an LLM agent to construct SQL feature programs and select a predictive model in a search phase, and then executes that program deterministically in the inference phase, making predictions based only on query-defined feature maps.

ClusterFusion addresses the difficulty of domain-specific text clustering by treating the LLM as the core clustering mechanism guided by lightweight embedding methods. This hybrid framework uses an LLM for topic summarization and assignment, allowing it to incorporate domain knowledge directly, which shows substantial gains in specialized domains compared to traditional methods.

The most significant work on the day is SIGS which introduces a neuro-symbolic framework for discovering analytical solutions to differential equations. This matters because finding exact solutions to coupled nonlinear PDEs is usually impossible without expert intuition or massive searching, and SIGS unifies symbolic reasoning with numerical optimization by using a context-free grammar to build valid blocks and then searching a continuous latent manifold for the best combination of those blocks.

This approach is powerful because it allows the system to recover analytical solutions for coupled nonlinear PDE systems, discover equivalent symbolic forms when the grammar is incomplete, and produce accurate approximations even when closed-form solutions are unknown SIGS. This framework improves existing symbolic methods by orders of magnitude in both accuracy and runtime across standard PDE benchmarks.

Another important piece is Do What You Know which presents an epistemic model for decentralized discrete-event systems using non-binary control. This work is crucial because it directly couples knowledge with action, formalizing the necessary and sufficient conditions for problem solvability in a way that mimics natural language, making it easier to revise requirements when initial conditions fail.

Moving down the list, MARS proposes a scalable knowledge graph question answering approach that uses large language models to link questions to knowledge graphs and iteratively retrieve relevant next-hop information without needing costly fine-tuning MARS. This method is valuable because it leverages the reasoning of LLMs while grounding their answers in explicit symbolic knowledge from KGs, offering a more predictable pipeline than purely agentic exploration.

We also saw some interesting work on how language models use group-linked fields, specifically Language Models Give Group-Linked Fields the Wrong Weight Language Models Give Group-Linked Fields the Wrong Weight. This paper investigates whether reliance on certain inputs in decision-making is warranted by evidence, finding that display format and predictive value matter more than just observed changes in reliance.

Finally, Learning When to Automate studies queue control in human-AI service systems using a UCB-DPP policy Learning When to Automate. This work is relevant because it models the trade-off between automation costs and human congestion in hybrid service systems, proving that the proposed policy achieves stability for the human queues while learning unknown system parameters.

The most significant contribution today involves developing PGL-Net, which addresses the real-world challenge of single image dehazing by creating a lightweight framework that uses physical inductive biases instead of complex parameter estimation AKMPC. This matters because current methods are either too fragile or too heavy for practical deployment where low latency is crucial. PGL-Net achieves state-of-the-art restoration quality while drastically cutting inference time, showing that physics can guide model design effectively in this domain.

This work builds on the general principle of using structured approaches to improve performance, much like how CAVE structures credit assignment for fragmented visual evidence CAVE. While CAVE focuses on reasoning paths, PGL-Net focuses on separating the dehazing problem into global distribution rectification and local structural refinement using modules like the Physics-Inspired Affine Fusion module to handle global alignment.

Another related area involves how we can structure learning processes for better results, which is seen in Vividh-ASR's findings on learning rates for speech recognition Vividh-ASR. This suggests that the timing of updates during fine-tuning can be tailored to the specific type of audio data, guiding parameter efficiency.

The concept of structuring reasoning is also explored in how agents solve complex problems, as seen in Deep Reasoning which uses structured meta-reasoning to build task-specific scaffolds Deep Reasoning. This contrasts with the physical guidance used in PGL-Net, but both aim to impose a beneficial structure on a complex process.

Finally, understanding the underlying geometry of learning is relevant because phase transitions in deep neural networks are governed by saddle points Phase transitions reveal hierarchical structure in deep neural networks. This provides a theoretical lens for why certain training dynamics occur, which could inform how we design the inductive biases used in models like PGL-Net.

The most significant development today concerns how we can truly make the latent spaces learned by generative models disentangled so that different semantic features are separated cleanly, which is crucial for controllable generation. This problem was addressed by showing that standard variational autoencoders and adversarial training methods consistently fail to enforce this desired factorization on the resulting posterior, leading to entangled representations. To combat this, researchers introduced non-parametric Maximum Mean Discrepancy or MMD to directly sculpt the aggregate posterior distribution, which allows for programmable priors.

This approach yielded promising results: it achieved a seventy percent reduction in detected entanglement compared to leading baselines and offered flexibility to inject semi-supervised information for better semantic alignment. Furthermore, they demonstrated precise shaping of the aggregate posterior distribution without sacrificing reconstruction quality through quantization-aware training. This work sets a new standard for disentanglement, and it builds upon prior efforts in representation learning like AKMPC.

Another important area involves making vision models more robust when applied to unseen data, particularly in tasks like human pose estimation where cross-domain generalization is key. A systematic analysis of features from different upsampling levels in the Stable Diffusion U-Net revealed which blocks provide the strongest robustness for pose estimation. This led to SDPose, which extracts features from these selected blocks and fuses them with a lightweight module to create a robust representation, optimizing keypoint heatmaps jointly with an auxiliary loss. This method achieved performance comparable to Sapiens on COCO and established the best zero-shot performance on cross-domain datasets, building upon prior work like SDPose.

The underlying mechanics of how internal representations form power-law anisotropy across various neural systems also received attention today. Researchers characterized this formation process by solving the learning dynamics of a two-layer linear neural network in a teacher-student setting with power-law input and teacher structures. They found that the local power-law exponent of the internal representation spectrum evolves nonmonotonically over training time, suggesting a general mechanism for how input statistics and task structure give rise to these representations. This finding connects to broader observations about internal dynamics, such as those detailed in Learning reshapes power-law anisotropy in internal representations.

For agentic systems, ensuring reliable multi-step workflows requires formal verification, which is a major hurdle. Lean4Agent proposes a framework using Lean4, a dependent-type formal language, to model and verify agent behavior semantically under explicit assumptions. Experiments showed that workflows verified by this method outperformed failing ones by an average of eleven point nine four percent, and subsequent work in Lean4Agent further improved performance on SWE-Bench tasks.

Finally, the reliability of large language models in translation tasks was explored through a retrieval-augmented generation system. This RAG+Prompt system integrated linguistic analysis and embedding-based retrieval to enhance Japanese-Chinese translation for sentences with noun-modifying clause constructions. The mean BLEU score increased significantly with larger knowledge bases, showing that this interpretable approach improves performance compared to baseline models.

The most important work here is the framework for sampling coherent but cognitively unavailable research directions, because it suggests a way to find scientific ideas that are plausible but outside the current community's reach, which is crucial for genuine discovery rather than just incremental progress. This method decomposes papers into idea atoms and scores combinations based on both scientific plausibility and community availability to maximize novelty.

This approach builds upon earlier work that shows how LLMs can generate novel ideas by recombining existing literature AKMPC — The Alien Space of Science: Sampling Coherent but Cognitively Unavailable Research Directions. Furthermore, the idea of using models to guide search and recommendations is relevant, as CORE optimizes content returned by generative engines to steer output rankings Controlling Output Rankings in LLM-Based Search via Generative Engine Optimization.

Another piece of work tackles the problem of monitoring reasoning traces in models, specifically showing that information theory is necessary but not sufficient for Chain-of-Thought monitors to be reliable Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory. This relates to how we can supervise complex reasoning processes, which is a key concern when exploring new scientific directions suggested by the sampling method.

We also see efforts to improve reasoning stability in models, such as the Halo Architecture which proposes moving from fuzzy floating-point arithmetic to exact rational arithmetic for better numerical stability From Fuzzy to Exact: The Halo Architecture for Infinite-Depth Reasoning via Rational Arithmetic. This foundational work addresses the underlying computational substrate that might limit what models can actually reason about.

Finally, there are methods for managing dynamic preferences in AI systems, like Adaptive Pluralistic Alignment which uses a jury system to track evolving values without needing constant retraining Adaptive Pluralistic Alignment: A pipeline for dynamic artificial democracy. This speaks to how we might manage the evolving consensus within a scientific field as new ideas emerge.

The GCA framework is the most significant development because it directly addresses the critical gap in climate decision-making: translating complex scientific evidence into actionable guidance for GCC states. This is important because general large language models are currently too weak for this region-specific task, and GCA Framework: A GCC Countries-Grounded Dataset and Agentic Pipeline for Climate Decision Support introduces a unified approach using a curated dataset and an agent to bridge this gap.

The core of the GCA framework involves building GCA-DS, which is a multimodal dataset containing two hundred thousand question--answer pairs covering policies, academic literature, and event reporting on heatwaves and floods. This dataset is coupled with remote-sensing inputs to link imagery with textual evidence. Building on this foundation, the Gulf Climate Agent orchestrates a modular tool pipeline that uses real-time signals and geospatial processing to create derived indices and visualizations.

This agent approach is supported by work like AKMPC, which shows that domain fine-tuning and tool integration substantially improve the reliability of LLMs on climate tasks compared to general-purpose baselines. Furthermore, the ability of models to handle novel information is being tested by studies like Disentangling generalization and memorization in large language models using chess, which found that model performance degrades sharply when the density of relevant priors decreases.

Another related area involves controlling model behavior at inference time; for instance, Conditional Optimal Bridge for Riemannian Activation Steering introduces Cobras to cast steering as a Schrödinger Bridge, which outperforms prior steering methods in controlling model outputs while largely preserving out-of-distribution performance. This contrasts with efforts to mitigate memorization, such as Output Vector Editing (OVE), where OVE-MA showed that adaptive position selection can concentrate suppression into fewer weight updates, leaving a measurable percentage of text recoverable.

Finally, the need for better multimodal understanding is highlighted by MMClima: A Framework for Multimodal Climate Science Data and Evaluation, which introduces a large-scale framework to benchmark multimodal language models on climate science questions across text, video transcripts, and figures. This work complements the GCA approach by providing a standardized way to evaluate how well models reason across these different data types in the climate domain.

Today's papers

The papers