AI papers — 2026-10-12

Today's focus is on developing a Koopman operator approach for controlling a two-spool turbofan engine because accurately modeling these complex dynamics is crucial for robust control under varying flight conditions. We are trying to build a single Koopman model that can be reused across different control strategies by adapting an extended dynamic mode decomposition with a cost function that captures both spool-speed dynamics and the engine pressure ratio. This identification method allows us to predict spool speeds and the engine pressure ratio accurately, which is the foundation for our adaptive Koopman-based model predictive controller.

This AKMPC then uses this identified model along with a disturbance observer to manage the system, offering better performance than a simple Koopman-based feedback linearization controller when flight conditions change. Furthermore, this framework lets us express nonlinear output limits, like rotor-acceleration and turbine-inlet-temperature constraints, as linear constraints within the AKMPC structure. This means we can enforce safety limits while maintaining good control.

The study showed that while all methods performed similarly at sea level, the AKMPC was superior when flight conditions varied because it handled nonlinearity and model mismatch better. Finally, controlling the engine pressure ratio using this strategy also resulted in an improved thrust response, showing the practical benefit of this approach for turbofan engine control.

The most important work here is figuring out if current video language models can actually grasp the flow of time in driving scenarios, because if they cannot, they are not reliable for safety. We tested various prompting techniques on these models for ego-motion tasks, and we found that standard methods like visual or chain-of-thought prompting are inconsistent across different models and tasks.

What really matters is that even though newer models show some improvement with few-shot learning, the core issue remains: lightweight LoRA finetuning combined with simple probes trained on intermediate visual features actually recovers strong motion discrimination. This suggests that the necessary motion cues are present within the model's internal representations, but the pathway to understanding that motion from video to language isn't being reliably utilized.

This finding is significant because it implies that these models are behaving more like frame-level recognizers driven by inherent biases rather than true temporal world models for driving. This contrasts with work on reward optimization, where GraftOOD shows that augmenting video reward signals with out-of-distribution exemplars can significantly boost dynamic degree in Wan2.1, raising it from 4.86 to 44.41 relative to a baseline method, even though the resulting motion isn't always perfectly plausible.

This external supervision approach is useful because it shows that adding examples helps refine the reward signal without necessarily equating higher measured dynamics with coherent, visually plausible motion in the generated videos. This idea of using external supervision complements other methods, such as those trying to detect drivable areas using thermal-RGB fusion, which offers a robust way to find wheel tracks on snowy roads by projecting future trajectories from GNSS data.

The work on CoCoEvolve is what matters most because it shows how meta-reflection can actually improve production AI systems, which is a huge step beyond just tweaking prompts or programs in isolation. This method works by having an agentic mutator look back at the entire evolutionary history to probe the live system before suggesting changes, which helps avoid getting stuck on isolated failures.

This approach has shown significant gains across several areas. For instance, when applied to data engineering workflows, CoCoEvolve lifted the test-set solve rate from fifty-nine point one percent up to seventy-seven point four percent, surpassing the best baseline of seventy-three point six percent and showing an eleven percentage point increase over GEPA on HotpotQA.

Furthermore, across six different LLM data pipeline workloads, CoCoEvolve outperformed every evaluated baseline. On a more specific level, when optimizing data agents and LLM-powered data pipelines, CoCoEvolve topped the public leaderboard on DABStep with ninety-two point eight percent held-out hard accuracy compared to seventy-eight point four percent for Meta-Harness. This demonstrates that the meta-reflection component is valuable for improving production AI systems.

The work on Unreliable Narrator Framing explores how framing explanations using unreliable narrator theory can be a cost-effective way to explain satirical news. It found that models are better at identifying satirical tropes in American sources than Australian ones, which results in a new dataset and benchmark. This contrasts with the focus on system optimization, as CoCoEvolve shows what happens when you actively guide the system's evolution rather than just letting it stumble forward.

The work on Sonic-AMG is particularly important because it tackles the complex problem of generating human motion in real time based solely on spatial audio. This is a crucial step toward more immersive and interactive spatial computing experiences. This framework encodes acoustic content and body-relative source geometry as separate tokens to condition a causal transformer that generates continuous motion latents.

The core challenge addressed here is the mismatch between training, where history is derived from ground truth motion, and inference, where generated motion must predict its own history. To bridge this gap, the authors introduced rehearsal training to combine full-sequence supervision with short autoregressive self-rollouts for prediction. They also proposed spatial feedback to ensure that the body-relative source locations remain consistent when generating motion during both training and inference.

This approach allows Sonic-AMG to achieve state of the art FID and R precision on SAM, meaning it produces high-quality motion samples. This capability extends beyond simple generation because it supports real-time online generation of coherent reactions across various sound events and source locations. This work connects to the broader field by showing how separating acoustic and spatial information can unlock novel generative capabilities for human interaction.

The connection between generative diffusion and optimal transport is crucial because it provides a mathematical lens for understanding how data distributions evolve during generation. It links noise processes directly to minimizing transport costs between states, suggesting that backward diffusion can be viewed as a gradual decrease in entropic regularization.

We also see how modes emerge during generation when comparing diffusion and flow matching with optimal transport. Specifically, the dynamical instability that separates these modes becomes an exponentially thin boundary layer at the start of generation, which is smoothed out by entropic regularization. This contrasts with other methods where this separation might be more pronounced.

Furthermore, for any generic forward dynamics with a confining potential, the backward process driven by the score satisfies a drift-based variational principle whose minimal cost is simply the difference in entropy between the starting and ending distributions. This idea of mapping diffusion to optimal transport relates to how different generative models handle mode transitions.

This is supported by work that shows how attention sinks behave differently depending on whether a model uses bidirectional diffusion or causal autoregressive structures, suggesting that the underlying mathematical principles governing these processes are interconnected across different architectures.

The work on domain alignment for cephalometric landmark detection is particularly important because it tackles how to make machine learning models generalize across different types of data while maintaining high accuracy in a critical clinical setting. This research proposed a strategy involving a regional facial extraction module and an X-ray artefact augmentation procedure. This resulted in the method ranking best in the MRE of 1.186mm and third in the 2mm SDR at the online validation leaderboard for the two thousand twenty-four CL-Detection MICCAI Challenge.

This success builds upon prior work that fine-tuned a segment anything model on a dataset of two thousand six hundred seventy nine annotated images from forty two buildings. This improved the mean Intersection over Union metric from forty nine point nine percent to sixty nine point six percent for multi class segmentation. This improvement in class segmentation is then used within group anything with radiance fields to better group building components in three dimensional environments, advancing the application of that framework.

Another line of inquiry focused on learning representations to guide pixel diffusion transformers, which revealed that simply having better representation quality does not automatically lead to greater generative utility because externally aligned semantics can be lost during synthesis. To address this, the representation guidance method was introduced. This method forms and exploits internal representations while using a low-dimensional bottleneck to guide the decoder generation by modifying the flow matching target.

This approach showed that RG-L/16 with four hundred fifty nine million parameters outperformed substantially larger pixel diffusion transformers on ImageNet two hundred fifty six by achieving an fidi of one point five five.

The theoretical analysis concerning infinite-width neural networks also yielded practical insights. It showed that the evolution of a multilayer perceptron under robust training is described by a linear first-order ordinary differential equation derived from a variant of the neural tangent kernel. This provides new techniques for finding appropriate hyperparameters and learning network weights in wide finite networks.

The work on mixtok addresses the challenge of efficient visual tokenization by introducing a hybrid tokenizer that combines a compact one dimensional global representation with sparse two dimensional tokens for local refinement. This method achieves highly competitive reconstruction performance, with an rfid of one point one nine five at an average budget of approximately one hundred twenty eight content tokens per image on ImageNet two hundred fifty six.

The most critical work here is SpireBench because it directly addresses the weakness in agentic large language models: their inability to reason strategically over long horizons. We introduced this benchmark by having models play the entire commercial deck-builder Slay the Spire, where early choices dictate long-term success. This matters because current benchmarks are too short and episodic, failing to capture how stored knowledge, reasoning, and luck interact over extended play.

The results show that while frontier models beat simple heuristic baselines and match average human performance in the first act of the game, they quickly fall behind as the run lengthens. What distinguishes strong models from weak ones is their ability to manage risk; strong agents preserve health rather than spending it on growth, whereas weaker ones misjudge risk over a run.

Furthermore, we found that models do not improve from memories they write across runs because those memories are often wrong and overly cautious, which suggests the performance relies heavily on prior game knowledge. This is complemented by the fact that our harness is the first to give a model the complete original game for comparison against existing setups.

This long-horizon challenge connects to other agentic work because understanding sustained autonomous work is key. Terminal agents are moving beyond simple local operations, and we introduced TerminalHorizon, an automated data synthesis engine designed to capture these complex patterns of exploration and strategy accumulation in real-world tasks. This engine constructs 1,500 task environments by using independent execution to identify missing key work and then uses that evidence to guide environment reconstruction.

The findings from SpireBench suggest a need for models that can maintain strategic coherence over extended sequences, which is a problem also touched upon when analyzing how models learn sustained behaviors. This contrasts with the work on Multi-Stream Attention Residuals, which introduced MSAR to improve performance in large-scale pre-training by allowing different heads to retrieve historical representations through distinct contextual views.

On a more foundational level, the work on Locally Private Parametric Methods for Change-Point Detection shows how statistical constraints affect localization accuracy. We established that the accuracy of finding a single change point is fundamentally limited by the Chernoff information between the pre- and post-change distributions. We designed algorithms based on randomized response to maximize this information under privacy constraints.

This statistical rigor contrasts with the empirical findings in speech processing, where we found that using automatic speech recognition confidence as a proxy for data quality led to 3 to 7 percent relative word error reduction over simple language identification filtering when pretraining Finnish SSL models. This shows how proxies can improve model performance despite data limitations.

Finally, the framework BLEND addresses the trade-off between personalization and generalization in federated vision-language models by introducing a loss function that blends global and personalized adapters with a novel anchor design to balance these two goals effectively.

The work that matters most is SOAR because it directly tackles evidence fabrication in search agents by explicitly separating external knowledge from internal knowledge streams. This separation, achieved through a Source-Aware Workflow, allows for precise credit assignment between these different types of information, which is crucial for grounding reasoning. This means the agent stops relying on unreliable internal shortcuts and starts using verifiable sources to make decisions.

This approach significantly improves performance on knowledge-intensive benchmarks by achieving a seven point three percent and five point four percent gain over existing baselines when tested on thirty billion and seventy billion parameter backbones, respectively. Furthermore, SOAR successfully reduces evidence fabrication by a substantial seventy-four point eight percent, confirming that making information source attribution explicit leads to more faithful reasoning.

The method also disentangles external evidence from internal knowledge into parallel streams using the Source-Aware Workflow before incorporating Source-Consistent Reward to guide policy evolution. This means the system learns how to use both external facts and its own knowledge in a structured way, which is a key step toward robust operation.

This structural improvement in reasoning is supported by other related work that focuses on improving model fidelity through better representation learning. CoVA, for instance, introduces a lightweight framework that composes Variational Autoencoder features online into stage-wise guidance for flow-matching training. This process aligns intermediate SiT features with the composed representations early in training, which helps yield semantically discriminative and spatially structured dense features.

This alignment is important because it corrects a hidden confound where averaging class embeddings attenuates their magnitude, which otherwise couples semantic interpolation with conditioning strength and can degrade recall while improving FID. CoVA achieves a significant reduction in SiT-XL/2 FID from seventeen point two to seven point four on ImageNet two hundred fifty six times two hundred fifty six, demonstrating its effectiveness in generating high-quality representations.

Moving toward evaluation, the generative augmentation pipeline aims to fix semantic coverage gaps in test and evaluation by using vision-language models for zero-shot binary labeling of semantic factors. This is followed by a diffusion-based image-to-image model that transforms existing covered images into underrepresented domains while preserving scene content. The final step involves statistical analysis of the augmented image embeddings to verify that the synthetic imagery actually improves dataset coverage before it is used to enrich the test set.

This augmentation process yields test sets that better distinguish models with and without exposure to missing semantics, leading to T&E conclusions that more closely mirror those from a fully covered dataset. This pipeline is validated on both image classification and object detection under day/night semantic gaps, showing its utility in creating more representative evaluation conditions.

Finally, CUBE addresses the challenge of open-set insect species recognition by modeling unknownness around each known species rather than as a single global unknown class. This framework learns class-conditioned boundary evidence from auxiliary species and uses them as shadow references during inference to allow for local rejection of visually confusable unknowns. When ecological metadata is available, CUBE-E adds a location-consistency term without altering the visual backbone, further improving performance on benchmarks by achieving high AUROC scores.

Today's papers

The papers

Important terms

Koopman operator approach
A method used to model complex system dynamics, like a turbofan engine, by transforming nonlinear systems into linear ones using an operator. This helps in creating robust control strategies for varying flight conditions.
Extended Dynamic Mode Decomposition (EDMD)
An identification technique used within the Koopman framework to predict spool speeds and engine pressure ratio. It's adapted with a cost function to capture both dynamics accurately.
Adaptive Koopman-based Model Predictive Controller (AKMPC)
A control strategy that uses the identified linear model and a disturbance observer. It performs better than simple controllers when flight conditions change because it handles nonlinearity and model mismatch well.
Meta-reflection (CoCoEvolve)
An agentic method where an AI agent looks back at its entire evolutionary history to probe a live system before making changes. This helps improve production AI systems by avoiding isolated failures.
Source-Aware Workflow (SOAR)
A technique that separates external knowledge from internal knowledge streams in search agents. This explicit separation allows for precise credit assignment, reducing evidence fabrication and improving reasoning.