Research papers — 2026-10-12

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

The research assistant is compiling the morning briefing by synthesizing the provided abstracts into a continuous article.

We are trying to get better feedback for fuzzing kernel code because current methods miss security issues hidden in argument values instead of just code paths. We are attempting to fix this by building an LLVM-based instrumentation framework that extends KCOV to extract data flow from function arguments and return values. This framework reports source-level parameters and composite objects using DWARF metadata, which gives fuzzers state-aware feedback for mutation guidance into value-dependent states.

This is important because it allows us to guide fuzzers toward interesting states based on the actual data being passed, not just where the code executes. This contrasts with older context-blind approaches that only look at edge coverage traces. Furthermore, security analysts will gain deterministic argument records for root-cause analysis without needing slow printk or kprobe overhead.

We are also looking at how to stabilize self-consistent diagrammatic approaches in quantum materials science because they often converge to unphysical fixed points in the most interesting parameter regions. The solution involves redefining the iterative procedure by inverting the unstable eigendirections of the Jacobian associated with that fixed point, which we showed successfully stabilizes physical fixed points across challenging regimes. This stability work complements our data-flow extraction efforts by showing how to rigorously handle complex system states.

The most important work right now is Any House Any Task because it tackles the massive challenge of making robots operate effectively in huge homes where instructions are vague. This research proposes a planner called AHAT that uses an LLM to turn messy human commands into concrete, solvable steps defined in PDDL, which are then solved symbolically to create long plans. This matters because current methods using just prompting break down quickly when the environment gets bigger or the instructions get confusing.

To make this planning more robust against ambiguity, they introduced TGPO, a reinforcement learning algorithm that mixes Group Relative Policy Optimization with external correction of intermediate reasoning traces. This means if the planner makes a bad move mid-plan, an external system can step in and fix it before the whole plan fails. This improved decomposition capability is crucial because it lets the system handle those complex, multi-step intentions humans give it in a real house setting.

The concept of handling varied sensory inputs is also gaining traction with CoRE-VLA, which creates a framework that routes action generation to different experts based on what sensors are actually available. This allows robots to gracefully degrade when they lose a sensor or when they have different types of sensors, making the policy much more reliable in messy real-world situations.

Furthermore, UniConFlow is making significant strides in motion planning by creating a unified framework that handles multiple constraints simultaneously, such as collision avoidance and actuation limits. They introduced a novel prescribed-time zeroing function to guide the trajectory generation process during inference, which helps the model adapt its output to different system models on the fly.

In terms of data for training, LLMail-Inject has created a massive benchmark of over two hundred thousand distinct prompt injection attacks, moving beyond small template sets. This dataset is important because it forces models to handle real-world adversarial attempts that involve complex chains of actions, showing that current defenses often fail when faced with diverse human writing.

Finally, the work on DiffMorphGrasp addresses the problem of transferring grasping policies between robots that have different physical hand shapes. By conditioning the diffusion process on structured representations of hand kinematics, they enable zero-shot generalization to unseen morphologies, which is key for scalable deployment across different robot hardware.

The work on DIDO is crucial because it tackles the latency issue in using video generation models for predicting future robot actions in manipulation by distilling multi-step predictions into a single step. This matters because iterative denoising, which is standard, adds delays to closed-loop control systems.

DIDO addresses this by combining distribution matching distillation with interaction-centric representation guidance to compress the dynamics of a multi-step video model into one forward pass. This means it explicitly models the gripper and manipulated object using supervised bounding box visual reasoning tokens and aligns the target object's representations with features from a pretrained DINOv3 encoder. This interaction-centric guidance helps the distilled model keep both relevant entities and their future dynamics intact in just one step, which significantly cuts down on inference time.

This distillation approach yields strong success rates across different benchmarks, achieving ninety-nine point zero percent on LIBERO and ninety-two point zero percent on RoboTwin. This performance shows that even with a single denoising step, the model retains the critical interaction dynamics needed for real-world robotic manipulation tasks. This method builds upon earlier findings where static background structure emerges early in the denoising process, while the interaction dynamics only appear later.

The most critical work right now is SAP-CoPE because it directly tackles how robots can navigate safely and comfortably in crowded indoor spaces by integrating human pose estimation with motion planning. This matters because current onboard sensing is too limited, and we need methods that respect human psychological comfort zones to generate truly social-aware trajectories.

SAP-CoPE introduces an uncertainty-aware optimization problem to estimate 3D human poses from camera observations, optionally using sparse LiDAR data while considering how the human joints move together. This pose information is then used within a model predictive control framework to build a personal space field, allowing the robot to plan for both physical safety and psychological comfort.

This planning module is supported by the perception module, which formulates an uncertainty-aware optimization problem that estimates 3D human poses from camera observations, incorporating optional LiDAR point-cloud measurements and considering human joint coherence. This estimation feeds into the MPC planner, which constructs a personal space field to guide the robot's movement.

The framework's effectiveness is shown through extensive real-world evaluations demonstrating its ability to generate social-aware trajectories for autonomous systems. This approach builds upon the perception module by using estimated human poses and orientation to construct that personal space field within the MPC framework.

The work that matters most is Staggered Integral Online Conformal Prediction because it provides long-run safety guarantees when synthesizing with safety-critical controllers like robust tube model predictive control, which is crucial for controlling uncertain systems. This method uses an integral score function to quantify the combined effect of disturbance and learning error, which means it gives a better picture of model uncertainty than standard online conformal prediction.

This approach is important because it allows us to maintain safety over time even when the system's dynamics are changing without direct measurements of state derivatives. It works by using an integral score function to lump together the effects of disturbance and learning error, giving a more robust measure of what is uncertain in the model. This is more significant than just having a prediction because it directly addresses the uncertainty in adaptive systems.

The validation for this method comes from simulating an all-layer deep neural network adaptive quadcopter using robust tube model predictive control, which shows that SI-OCP can handle complex learning parameterizations and control strategies in practice. This simulation demonstrates the practical applicability of the algorithm to real-world, complex control scenarios.

Sampling-Based Safe Reinforcement Learning is also significant because it tackles the fundamental challenge of safe exploration in reinforcement learning by enforcing constraints directly across a finite set of dynamics samples. By doing this, it approximates an otherwise intractable worst-case optimization over uncertain dynamics, which gives practical safety guarantees in continuous control domains. This formulation also introduces an exploration strategy that constrains epistemic uncertainty instead of relying on explicit bonuses.

This is more important than the specific exploration strategy because the core idea is maintaining safety throughout the entire learning process, not just finding a good policy eventually. The algorithm achieves high-probability safety guarantees and a finite-time sample complexity bound for recovering a near-optimal policy under regularity conditions.

Featurized Occupation Measures are also noteworthy because they offer a way to handle the curse of dimensionality in numerical optimal control by shifting the difficulty from the state dimension to the interconnection topology. This finite-dimensional interface allows HJB subsolutions to guide primal search while occupation data refines certificates, which is useful when dealing with complex systems.

This is less significant than handling high dimensions because it deals with a more fundamental problem in numerical methods: how to make global searches tractable when the state space is too large. The method proves that these certificates remain reusable even under bounded model perturbations, which adds robustness to the solution.

X Hacking addresses a critical issue in explainable AI by showing how easily analysts can manipulate XAI metrics like SHAP values to support desired conclusions, which undermines trust in model predictions. This vulnerability is particularly concerning because it shows that exploiting model multiplicity at scale can be done efficiently through Bayesian optimization.

This is less significant than the threat of manipulation because it deals with the integrity of the explanation itself rather than system safety or core control guarantees. The paper suggests detection methods and ethical considerations, which are important but secondary to immediate operational concerns.

The work on Arbitrum's Timeboost mechanism is most important because it directly addresses how decentralized finance applications handle transaction ordering and potential manipulation through maximal extractable value or MEV. This mechanism attempts to create an express lane for transactions, but the empirical study shows that control is highly centralized, with just three entities capturing nearly all of the auctions.

While this access does provide earlier inclusion for arbitrage activities on CEX-DEX platforms, its overall utility is limited because about thirty percent of timeboosted transactions still revert to their original state, meaning it doesn't effectively stop spam or other MEV inefficiencies. Furthermore, the secondary markets for reselling these express lane rights are struggling because they lack reliable execution and sustainable economics. This decline in auction competition is also linked to the introduction of dedicated execution lanes by centralized exchanges that allow for pay-per-transaction billing instead of round-based fees, which has reduced revenue for the Arbitrum DAO.

The fairness-guided deep Q network approach is what really matters here because it directly addresses severe coexistence imbalance when new radio technology shares spectrum with existing Wi-Fi systems. The work tried to adjust the transmission opportunity time, or TXOP, for the new radio based on how much throughput each node was actually getting.

One reward design, Q1, focused on keeping the per-node throughput perfectly balanced across all radios. This strict approach showed a clear cost in terms of overall efficiency and fairness recovery. Conversely, Q2 allowed for a wider region where fairness was less strict but rewarded improvements when things were in an intermediate state of balance.

The comparison against existing methods shows that this new DQN policy performs well across four different priority classes and asymmetric user populations when measured by aggregate throughput and Jain fairness metrics over twenty trials. This means the system can handle real-world deployment scenarios better than the default fixed TXOP baseline or even the standard listen-before-talk method.

This result is significant because it characterizes exactly how much efficiency you lose when you enforce strict fairness, while also showing how much performance you can recover by relaxing those fairness constraints under moderate conditions. This contrasts with the work on managing risk in defense systems, which deals with a different kind of constraint entirely.

The work concerning preference-conditioned multi-objective reinforcement learning is most significant because it tackles a core problem in applying reinforcement learning to real-world systems like heavy-duty vehicles where safety, efficiency, and cost are all competing priorities. The framework proposes using generalized policy improvement and proximal policy optimization to learn policies that explicitly capture trade-offs between these three objectives across a simulated highway traffic environment. This approach yields smooth and interpretable Pareto fronts, meaning the learned policies clearly show how different driving behaviors balance these conflicting goals.

This method allows for transitions between different driving behaviors without needing to retrain the entire system, which is a major practical advantage for adaptive decision-making in autonomous trucks. The evaluation showed that these learned policies capture meaningful trade-offs when compared against an analytical baseline across various traffic densities. This capability builds upon the idea of using multi-objective reinforcement learning to manage complex operational requirements.

Another area of research delves into the microwave response of superconductors with paramagnetic impurities, aiming to predict how these impurities affect device performance in superconducting electronics. Theoretical methods connect the sub-gap quasiparticle spectrum generated by pair-breaking to the frequency and temperature dependence of conductivity for superconductors operating at microwave frequencies. These calculations predict anomalies in dissipation, quality factor, and resonant frequency for high-Q SRF cavities sensitive to non-magnetic and paramagnetic disorder at low temperatures.

This theoretical work suggests that understanding impurity effects is crucial for designing superconducting devices where performance metrics like dissipation are highly sensitive to material purity. This contrasts with the machine learning approach by focusing on fundamental material physics rather than adaptive control strategies.

In a separate domain, research into unconventional superconductivity in ScIr two-x-Si-x family chiral crystals reveals evidence of superconductivity driven by strong electronic correlations stemming from flat bands near the Fermi level. The study uses muon-spin spectroscopy combined with band structure calculations to describe the pairing using a two-gap model, indicating an unconventional nature due to nodal gaps.

This material science finding is distinct from the control and electronics work, focusing instead on discovering new states of matter in complex crystalline structures. This challenges conventional models of superconductivity by linking it to topological features and strong electronic interactions.

Meanwhile, efforts are being made to develop low-cost prosthetic hands using electromyography and machine learning to improve accessibility for amputees in developing countries. The system classifies five different muscle gestures using a shallow neural network trained on time domain EMG features, achieving high classification accuracy. This classification then controls a robotic hand with two degrees of freedom, adding functionality through a wrist rotation mechanism.

This biomedical application focuses on signal processing and machine learning to create tangible assistive technology for human augmentation. It is an engineering solution aimed at practical, low-cost deployment rather than theoretical physics or complex control theory.

Today's papers

The papers

Important terms

LLM-based instrumentation framework
A new system built using LLMs to extract data flow from function arguments and return values in kernel code. This helps fuzzers guide mutations toward states based on actual input data, not just code paths.
AHAT (Any House Any Task)
A planner that uses an LLM to convert vague human commands into concrete, solvable steps defined in PDDL. It's designed to make robots operate effectively in large homes with confusing instructions.
TGPO
A reinforcement learning algorithm that mixes Group Relative Policy Optimization with external correction of intermediate reasoning traces. This helps planners fix bad moves mid-plan for better robustness.
SAP-CoPE
A framework integrating human pose estimation with motion planning to create a personal space field for robots. This allows them to navigate crowded indoor spaces safely and comfortably by respecting psychological comfort zones.