AI papers — 2026-10-20

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 how we can better evaluate uncertainty when segmenting medical images because looking at error scores does not show if the uncertainty estimation helps a doctor make a better decision. The goal is to move beyond simple error maps to create a protocol that looks at how uncertainty estimation, the policy for deferral, and the review budget work together.

The most significant piece of work in this area is Saket Maganti Shreyank N Gowda's proposal on A Decision-Aware Evaluation Protocol for Uncertainty in Medical Image Segmentation AKMPC. This protocol tests different setups, such as using test-time augmentation versus probability-derived scores, to see how they affect error reduction and risk coverage across tasks like retinal vessel segmentation.

Another important development is the study of how vision-language models learn to explore in 3D space, detailed in Learning to Think and Explore in 3D World ACT3D. This work shows that using reinforcement learning driven by final answer correctness allows a model to learn an effective exploration policy. This policy guides the model to acquire necessary viewpoints for complex reasoning, leading to significant accuracy gains over its base model.

We also have research into how language models can improve text-to-image generation through prompt engineering, specifically SCOPE SCOPE. This method uses reinforcement learning to rewrite prompts based on semantic structure within the image generation process. It shows that structuring the prompt semantically helps guide the model better than simple single signals.

Finally, we see progress in how models can leverage intermediate representations for few-shot adaptation with LAMP LAMP. This framework combines global and regional information from different encoder layers without training. Mixing these features improves classification accuracy when only a few labeled examples are available.

The most pressing issue now is understanding how large language models behave when they are supposed to provide mental health support because demand is growing faster than human clinicians. If we do not know what therapeutic interactions look like, we cannot trust them with vulnerable users.

One key effort involves mapping out the conversational strategies used by these models. The work by AKMPC introduces a new way to categorize ten distinct therapeutic moves that models favor or ignore. These models tend to over-use probing and neglect certain cognitive behavioral therapy techniques, suggesting we can bring them closer to human clinical standards by exposing these categories as callable tools.

This idea of structured interaction relates to how models process information in general. For instance, the work on Hear2Act shows that prosodic cues are only useful when the user's need is missing from the text. This reveals a gap where making information explicit in text leads to much better decisions than relying on raw audio.

Another area of research touches on how models learn from data, even when that data is intentionally flawed. The paper When Clean Data Hurts demonstrates that optimal binary learners are surprisingly vulnerable to adversarial corruption. Monotone adversaries can make learning a multiclass problem completely unlearnable under certain conditions.

Moving into more complex reasoning, the agentic approach in MedPAO-VL shows how an agent can iteratively refine a structured report by observing semantic discrepancies. It achieves high expert ratings by prioritizing explainability over monolithic generation, mirroring the sequential decision-making process that the therapy ontology in AKMPC seeks to map out for LLMs.

The most significant development is the introduction of FeaTok, which addresses a fundamental problem in discrete image generation by shifting the training objective from pixel reconstruction to reconstructing pretrained features. Standard tokenization methods often prioritize low-level fidelity, which can hinder a generator's ability to model the underlying structure needed for high-quality output.

By training tokens on features instead of pixels, FeaTok has shown improvements across various models like SigLIP and DINOv2, leading to up to a twenty-nine percent reduction in gFDr6 scores at 256 by 256. This feature-loss tokenization approach is also highly effective in class-conditional generation.

It reduced gFDr6 by up to twenty-nine percent over the strongest discrete baselines and achieved three times less sampling compute compared to the best unguided continuous baseline, which is a major efficiency gain for practical use. Furthermore, FeaTok has surpassed prior art in human preference reward-model scores without needing reward-based training.

The work on OpenGCUA is important because it moves the field toward creating fully open stacks for generalist computer use. This involves defining a generalist computer-use agent that can choose how to observe and act across visual and programmatic interfaces. Frontier systems like GPT-6 Astra's computer-use system remain closed, limiting systematic study of their design. OpenGCUA introduces code execution as a key feature, allowing intermediate results to determine subsequent operations, which is a significant step toward more transparent agent design.

This idea of integrating model capabilities into interaction is supported by the work on Perceive, Predict, and Plan: a Unified World Action Model for Egocentric Dexterous Manipulation. This model integrates 3D perception, action-conditioned prediction, and goal-directed planning through three modes: Perceive to reconstruct motion, Predict to generate futures under actions, and Plan to generate trajectories with anticipated consequences. The unified training approach has been shown to improve all three capabilities over single-mode training.

Finally, research into routing policy optimization for online diffusion reinforcement learning offers a more efficient way to align visual generators with reward signals without introducing stochasticity through trajectory perturbation. Routing Policy Optimization turns attention routing into a stochastic exploration action, sampling routes based on the generator's native attention mass while using a likelihood score-function objective. This method improves generation quality while achieving nearly four times higher training efficiency than Flow-GRPO-style methods and restores the original deterministic sampler at inference time.

The most significant development today involves a new method for making spiking neural networks more energy efficient by introducing a state-driven mechanism for neuron dormancy. This directly tackles the persistent computational overhead caused by the continuous subthreshold dynamics of postsynaptic neurons, which currently burdens up to forty percent of neurons in models like ResNet-18, even when they are not firing.

The Dormant Spiking Neuron model introduces a state-driven dynamic gate that allows a neuron to enter a dormant state when its membrane potential drops below a certain negative threshold. This effectively prunes redundant computations at the instance level. This dormancy mechanism is trained end-to-end using backpropagation through time with surrogate gradients, allowing the network to learn an adaptive dormancy policy. It cuts energy consumption by forty percent to over seventy percent across various datasets.

This approach builds upon earlier work that focused on improving visual perception in multimodal large language models. The paper Seeing is Believing: Benchmarking and Mitigating the Interference of Language Priors on Visual Perception in MLLMs introduces ViSiP-Bench, a new benchmark designed to measure prior resistance more precisely than just accuracy alone. This suggests that simply scaling up models does not improve prior adherence.

This leads to the proposal of MAPS, which fine-tunes models on factually grounded knowledge to reduce excessive reliance on priors without losing general capability. In terms of reasoning capabilities for multimodal agents, MMSearch-OPD presents a new framework that combines controllable data refinement with joint reinforcement learning and on-policy distillation.

This method refines reasoning chains by expanding them based on existing simple multimodal questions. It uses parameterized transformations to control search depth and cross-modal evidence requirements while guiding the agent's behavior through reward gaps modulated by token-level teacher guidance.

For visual world modeling, Velox addresses the bottleneck in chunked video diffusion models where context retention slows down as rollouts get longer. Velox uses a training-free attention router that represents historical blocks as camera frustum slabs to modulate layer importance based on geometric context. This accelerates attention by forty-three percent at chunk forty-eight.

DK-VLM proposes Dual-Kernel Attention for vision language models to better handle 3D spatial relations across views. It computes attention from both a visual token's input position and its geometric position in the scene using separate semantic and geometric kernels. This allows geometry to enter attention as position rather than just fused reconstruction features.

Finally, BEAR-Bench establishes a new benchmark for evaluating multimodal models on text-dense, professional documents in both English and Russian. This resource is crucial because it moves beyond simple information extraction benchmarks by testing how well models reason about complex documents in a bilingual setting.

The work that matters most is the development of Multi-Source Consensus Distillation which aims to make fine-tuning more robust against unwanted behaviors like poisoning or subliminal learning by learning from agreement across multiple datasets. This approach, MSCD, involves fine-tuning one model on each dataset and then using the consensus among these models to create new training responses for a fresh student model.

This distillation method can suppress undesirable behaviors while keeping the intended capabilities intact in practical tasks like medical advice generation AKMPC. The idea of using consensus to reduce reliance on any single source is a key mechanism here, and this is supported by the findings in AKMPC.

Another important area is improving factual accuracy in agentic systems through Iterative Nugget Optimization, which refines knowledge base entries iteratively until they are discoverable by the production agent INO. This method focuses on actionable factual corrections rather than generic feedback and has shown consistent improvements in terms of discoverability and usage of these corrections across various B2B agents.

In the realm of model architecture, the work on Homogeneous Vector Capsules introduces a new way to parameterize capsules that allows for adaptive gradient descent. This can improve classification accuracy without increasing model size Homogeneous Vector Capsules Enable Adaptive Gradient Descent in Convolutional Neural Networks. This technique demonstrated improvements in top-1 and top-5 classification accuracy on the ImageNet ILSVRC 2012 dataset for both simple monolithic CNNs and Inception v3 architectures.

Moving into video models, Token-Budget Distillation addresses the high computational cost of adapting video vision-language models by using a dual-path teacher student design with FlashVID compression Token-Budget Distillation. This framework successfully preserves full-token semantics under aggressive token reduction, showing strong performance retention on models like LLaVA-Video.

Finally, research into speculative forcing shows how to speed up autoregressive video generation by treating intermediate clean estimates as drafts Speculative Forcing for Efficient Autoregressive Video Generation. This technique reduces computation while maintaining the quality of full-step sampling in various video generation scenarios.

The most significant piece of work is EKLAV because it tackles how to properly train models to learn complex reasoning steps rather than just memorizing answers. It shows that supervising only the final answer can lead to better reasoning traces when placed correctly in the prompt, improving performance on BRIGHT passage reranking and showing superior results on table-reranking benchmarks compared to other distillation methods EKLAV.

ZipSplat is important because it offers a fast way to reconstruct 3D scenes from images using feed-forward methods, which is a major step toward real-time 3D understanding. It introduces ZipSplat by predicting a compact, observation-grounded scene representation whose size is set at inference. This grounds tokens in observations and uses geometric losses like the cone loss to ensure Gaussians stay near their source patches.

This approach outperforms pixel-aligned and token-based baselines on several benchmarks. EngTrace matters because it provides a symbolic benchmark for evaluating engineering reasoning processes, moving beyond just checking final answers which can be misleading. It generates gold reasoning traces with all intermediate quantities, allowing an evaluator to check coverage of these milestones, revealing differences in derivation that final answers often miss across various LLMs.

Bridging the Markerless Tracking Validation Gap is important because it creates a clinically viable pipeline for validating markerless intraoperative tracking of spinal anatomy using routine fluoroscopy data. This framework links pose recovery from multi-view CT to fluoroscopy with RGB-D frames and its ex vivo validation against ovine vertebrae confirms accuracy sufficient to serve as independent ground truth for other markerless methods.

Seeing Both the Forest and the Trees is significant because STRUCTPARTS addresses the challenge of generating 3D assets that possess both fine geometric detail within parts and structural coherence across the whole object. It achieves this through a dual-frame coordinate interleaving mechanism that allows shape latents to be interpreted under both part-local and object-global references, leveraging a segmented mesh scaffold for training.

Flow-CLIP is noteworthy because it refines global image-text alignment by incorporating visual patch context into token queries. This improves compositional matching accuracy over standard CLIP by using Image-Conditioned Evidence Refinement and Cross-Image Margin Ranking to encourage better binding between language and vision.

Retrofitting Recurrent Vision-Language Models from Pretrained Backbones is important because it allows the reduction of the parameter footprint of pretrained VLMs by adapting depth recurrence into the language backbone. This leads to a model that outperforms unpruned baselines on reasoning-oriented benchmarks while storing fewer parameters.

SCCL-DT is relevant as it proposes Semantic-Conflict-Aware Contrastive Learning to improve zero-shot chest X-ray diagnosis by tracking modality-specific similarity distributions and using dual teacher rules to consolidate rankings. It achieves a mean zero shot AUC of 0.850 on public benchmarks.

The most significant work here is Hybrid Latent Attention for Looped Language Models because it directly addresses a major performance bottleneck in modern large language models by improving decoding speed while maintaining high accuracy. This method keeps only recent tokens exactly and compresses older ones into compact latent representations that the query reads directly, which shrinks the cache size significantly.

This leads to substantial gains in throughput, with improvements up to seven times at longer contexts, and it keeps over ninety-seven percent of the original accuracy on complex reasoning benchmarks. This improvement is built upon a foundation established by VeoPlace, which uses vision-language models to guide chip floorplanning through an evolutionary search strategy based on visual proposals.

VeoPlace shows that VLMs can effectively complement traditional placement algorithms, boosting performance by over thirty percent on several benchmarks. Furthermore, MJev-Spatial introduces a System 1 multimodal decision model that tackles spatial understanding tasks by treating candidates as classes and scoring them using pretrained projections from a backbone. It achieves the highest overall score among tested models.

Another area of progress involves improving robustness in perception; VesperMUSE creates a unified framework for image semantic segmentation that handles diverse lighting conditions by integrating RGB features with thermal representations and surface normal maps. This allows the network to selectively attend to relevant cross-modal cues, which is critical for autonomous systems operating in unstructured environments.

In the domain of activity recognition, Truly Rotation-Invariant HAR addresses a deployment shift where IMUs are reattached across different body locations. This makes robustness a structural property rather than an added augmentation. This framework reshapes sensor streams into triaxial vectors and fuses invariant features to classify activities reliably under independent per-location orientation shifts.

Finally, there is ongoing research into how skills transfer between models; Break It Down, Pass It On investigates whether inducing skills at the subtask level or using text formats leads to more reliable skill transfer across different tasks. The authors propose a skill utility score that combines specificity and abstractness to diagnose this memory effectively.

The most significant finding relates to how non-Gaussianity in language model activations offers a direct signal for finding interpretable structure. This is important because understanding these directions allows us to better control model behavior. The work that matters most, Non-Gaussianity Reveals Interpretable Directions in Language Models, shows that projection distributions of residual-stream activations along sparse autoencoder decoder directions are substantially more non-Gaussian than those along random directions.

This suggests that interpretable directions possess a specific statistical signature. This signature is then exploited by the workflow introduced in Non-Gaussianity Reveals Interpretable Directions in Language Models, which is called ICA Lens. This workflow efficiently discovers, annotates, and evaluates these non-Gaussian directions across models like GPT-2 Small and Gemma 2 2B.

The results indicate that the recovered directions capture coherent lexical, syntactic, and semantic patterns under both human and LLM evaluation. The significance of this discovery is further supported by how non-Gaussianity correlates with interpretability in Non-Gaussianity Reveals Interpretable Directions in Language Models, showing a positive association with both human-label confidence and automated interpretability. Furthermore, the resulting coordinates from this method are competitive with sparse autoencoders across several tasks while using significantly less memory and time than traditional methods.

Other related work builds on these ideas in different domains. Let the Batch Decide, On- or Off-Policy: Effective Sample Size for RL Post-Training addresses fragility in reinforcement learning when using nominally on-policy methods by proposing P3O, a batch-adaptive objective that replaces fixed clipping with a normalized effective sample size of per-token policy ratios. This approach helps keep updates close to the on-policy update when data is consistent.

In vision and language tasks, Q-Reflector proposes learning GUI action reflection using Qwen3-VL-8B, achieving high accuracy in assessing whether executed actions were valid steps toward a goal. This work shows how models can learn to reflect on their own execution success in complex environments.

Finally, FusionAudit: Pathway-Conditioned Robustness Auditing for Native Multimodal Models introduces PaRMA, an evidence-based auditing framework that uses pathway-specific diagnostics to expose architecture-specific blind spots in vision-language models. This method moves beyond simple pixel perturbation to uncover deeper vulnerabilities related to how native models attend to different types of visual information.

Today's papers

The papers

Important terms

Decision-Aware Evaluation Protocol
A new framework for testing how uncertainty estimation, deferral policies, and review budgets work together to help doctors make better decisions in medical image segmentation.
Reinforcement Learning for Exploration
Using reinforcement learning driven by the final answer correctness allows vision models to learn effective exploration policies that guide them to acquire necessary viewpoints for complex reasoning.
FeaTok
A method for discrete image generation that shifts the training objective from pixel reconstruction to reconstructing pretrained features, leading to significant quality and efficiency gains.
Non-Gaussianity Reveals Interpretable Directions
This technique finds specific statistical signatures in language model activations that correspond to interpretable directions, helping researchers understand and control model behavior.