Research papers — 2026-10-07

Today's work centers on bridging the production gap for quantum-safe cryptography using a hybrid-by-default Python library called quantum-safe. This matters because while the core algorithms for post-quantum cryptography have been standardized by FIPS 203, 204, and 205, the practical tools needed to implement them safely are still missing. The library scores itself against nine external requirements from bodies like CNSA 2.0 and CMVP to ensure it meets necessary standards.

The audit of nine PQC libraries revealed that hybrid combiners are present in only three out of six examined libraries, which is less than the one found in a March-two-thousand-twenty-six snapshot. Furthermore, discovery and inventory tooling remains a significant gap because none of the seven audited libraries possess it, so quantum-safe adds a CycloneDX CBOM to address this.

Regarding specific implementations, only Bouncy Castle holds a CMVP FIPS 140-3 certificate. quantum-safe reports two hundred twenty-five runnable known-answer cases as conformance evidence. However, the library's defaults fall below the CNSA 2.0 parameter sets when scored against those same requirements.

In terms of performance, a full X25519 plus ML-KEM-seven hundred handshake takes about two hundred thirty-one microseconds under Docker and Linux. This is five point three times slower than an X25519-only handshake but only about two point three percent of the typical TLS one point three budget.

The dataset HRDexDB provides a real-world four dimensional grasping dataset capturing three thousand two hundred trials across five embodiments, including human and robot hands interacting with objects. This resource is useful for studying embodiment-dependent grasp strategies through human-to-robot contact map transfer and visual robot-object contact estimation.

Finally, the Deep Time-Series Forecasting survey introduces a taxonomy that jointly covers both backbone architectures and loss functions to model autocorrelation in history sequences. This helps structure the field by providing a holistic overview of how these two aspects evolve together.

The most important finding here concerns the limitations of current large language models when dealing with non-textual sensory data, specifically audio-visual content. The AVMeme Exam benchmark showed that state-of-the-art multimodal models struggle significantly when analyzing textless music and sound effects, performing poorly on context and culture compared to surface content. This suggests a major gap in how these models perceive meaning beyond what is explicitly written or seen.

This difficulty stems from the nature of the data itself; internet audio-visual clips convey meaning through time-varying sound and motion that text cannot capture alone. The benchmark involved over one thousand iconic sounds and videos, each paired with questions assessing understanding from surface content up to world knowledge. While models handled the explicit parts well, they failed when asked to interpret cultural nuances embedded in music or sound effects.

This finding is crucial because it calls for a shift toward multimodal intelligence that can genuinely perceive contextually and culturally rather than just processing surface-level inputs. This limitation contrasts with other areas of research, such as the work on SubMAPG, which uses submodular team utilities to improve online task allocation in multi-agent systems by developing novel policy gradient estimators.

Another area where models are being refined is in agent security, as MiniScope proposes a task-centric permission model that automates least-privilege enforcement for AI agents. This significantly reduces the need for manual permission confirmations and addresses the usability versus security tradeoff inherent in existing flat permission structures.

Similarly, interactive imitation learning in robotics is trying to build more robust behavior by allowing robots to improve their actions based on intermittent human feedback during execution. However, this field currently suffers from a lack of unified terminology and structure across different formulations, which slows down progress.

The work on self-sustained Josephson dynamics in dipolar supersolids matters because it provides a theoretical pathway to understanding how these complex quantum systems maintain coherent oscillations. This is crucial for designing novel quantum devices. The authors used an asymmetric two-mode model to describe a triangular dipolar supersolid and found that the system's symmetry allows for a consistent two-mode mapping even with seven droplets present.

This mapping lets them straightforwardly determine the self-trapping regime of the associated Hamiltonian. This determination is significant because it links the abstract mathematical description directly to observable physical states within these supersolids.

Furthermore, bringing this system into rotation preserves its capacity to sustain Josephson junction dynamics across its entire operational range. They assessed how well their asymmetric two-mode model holds up under these rotational conditions. The off-axis droplets move radially during evolution in a way that aligns with the size of the central droplet, and this movement does not interfere with their model predictions.

In parallel, research into training trustworthy web agents is addressing a critical gap where synthetic environments often fail because they only look plausible but lack actual backend consistency. The framework constructed here treats each generated website as a structured scaffold containing pages, links, and database records to ensure it is executable and auditable before policy training begins.

This verification process improves the feasible-task rate from forty-eight point six percent to ninety-four point eight percent across five hundred synthetic environments spanning six domains. This success shows that verified synthetic environments can serve as a reliable substrate for compact web agents, shifting the focus from surface plausibility to state-grounded supervision.

This contrasts with other agent training methods that might rely on less rigorous environment generation. Meanwhile, work on quantum information explores how non-orthogonal states can be distinguished using a two-reservoir quantum machine driven by this discrimination process.

This machine maps its functional boundaries across the parameter space defined by the state overlap and Carnot efficiency, revealing phase transitions between pure heat-engine phases and dissipative phases. They identified critical thresholds where strong thermal driving guarantees positive work extraction, while weak driving leads to an anomalous reentrant transition that restores engine functionality after a purely dissipative regime.

The work on SchemaGraphSQL is particularly important because it directly addresses a major bottleneck in making large language models useful for querying complex databases. By using pathfinding graph algorithms to link schema elements, this method allows models to generate much more accurate SQL queries, which is vital for practical applications.

This approach works by first building a schema graph from foreign key relations and then using a single prompt with Gemini 2.5 Flash to find the correct source and destination tables in the user's question. Before applying pathfinding algorithms, it determines the best sequence of joins. This entire process is designed to be zero-shot and training-free, making it cost-effective and scalable compared to previous specialized methods.

The core mechanism relies on classical pathfinding algorithms after the initial table identification step. Once the model suggests which tables are involved, a graph algorithm precisely identifies the optimal join order for the LLM to use when writing the final query. This is followed by post-processing steps to refine this sequence, ensuring maximum accuracy before the SQL is generated.

The results show that this framework achieves state-of-the-art performance on the BIRD benchmark, meaning it performs better than existing complex, multi-step LLM approaches that require extensive fine-tuning. Ablation studies were also conducted to look closely at how precision and recall trade off within the proposed framework. Furthermore, evaluations of the schema filtering method compared against other techniques across different model sizes provided deeper insight into its practical utility.

The most critical work right now is PhysCaP because it addresses a fundamental limitation in how robotic agents learn to interact with the physical world. This moves beyond just imitating demonstrations to actually understanding the underlying physics needed for manipulation. This means we can build robots that can handle novel physical scenarios without needing endless pre-programmed examples.

PhysCaP introduces a way for code-as-policy agents to seek out information through active interaction rather than just passively observing things. Vision-language policies often fail when the necessary physical properties, like an object's mass or stiffness, are not explicitly shown in the training data. The system tackles this by having a planner decide when to explore and a prioritizer that filters out interactions that seem physically impossible, which keeps the exploration efficient.

The core mechanism involves training-free modules that estimate these physical properties directly from the robot's own movement feedback, such as proprioception. This estimation is then used to guide the agent's decision-making process in real-world tasks and simulations like LIBERO. This capability means PhysCaP can achieve performance comparable to existing methods while requiring fewer interactions and less time to complete a task.

This approach builds upon the idea of representation quality, similar to what VLAct does for vision-language models, but instead focuses on grounding that representation in tangible physical constraints. The success in achieving high success rates on platforms like RoboCasa-GR1 using only a fraction of the full data highlights how focusing on physics-informed exploration can be a powerful axis for embodied AI progress.

The work on trace-class spectra of irreducible Gaussian quantum Markov semigroups is crucial because it establishes the fundamental spectral properties governing how quantum systems evolve under certain probabilistic rules. The drift matrix dictates this evolution, and when the linear flow generated by this matrix is not periodic, we find the spectrum settles into a closed left half-plane.

This classification is further refined by checking the real parts of all drift eigenvalues; if every single one has a negative real part, then every interior point in that plane represents an eigenvalue. However, if any drift eigenvalue possesses a positive real part or if all of them have zero real part, we find that there are no trace-class eigenvectors at all.

When the flow is periodic, the spectrum takes on a more complex shape consisting of unions of horizontal half-lines or filled parabolic regions. These results show that irreducibility cannot generally be ignored when studying these systems, and this framework can be applied to concrete examples like oscillators and boson chains.

The dOPT framework addresses a significant challenge in optimization by allowing the differentiation through embedded conic problems without needing to differentiate the entire formulation directly. It achieves this by reducing the problem at a computed primal-dual solution to an equality-constrained quadratic program that keeps the reference solution and its first-order sensitivity intact.

This reduction captures the local first- and second-order geometry relevant for differentiation, which is particularly useful when dealing with convex NLPs, QPs, SOCPs, and SDPs. Numerical tests confirm that computing the gradients using this method requires only a single symmetric linear solve. This makes it scalable and faster than existing differentiable conic optimization techniques as the problem size grows.

The MemHint pipeline tackles the persistent issue of memory leaks in C/C++ software by combining large language models with symbolic reasoning tools like Z3. It uses an LLM to classify functions as memory allocators or deallocators, then validates these summaries against the control-flow graph to ensure claimed memory operations are reachable on any feasible path.

This approach successfully detected fifty-four unique memory leaks across over three million lines of code in real projects, outperforming other static analysis tools significantly. This pipeline is then integrated into CodeQL and Infer through specific extension mechanisms, allowing the validated summaries to be used for bug detection.

The phase-space geometric measure of magic in qubit systems provides a clean characterization of quantum computational advantage beyond stabilizer circuits using the l 1 distance from a state's discrete Wigner function to the stabilizer polytope. The study proves that for single qubits, this metric corresponds exactly to a cuboctahedral unit ball and determines the maximum value attained at eight specific face states.

The geometry also dictates that for two qubits, enumerating all translation-covariant frames reveals four equal classes where the ratio of metrics takes on values of one half, one whole, or two. Within a specific frame called Wootters' frame, a tetrahedral dichotomy governs when the product bound is saturated based on integer tightness ratios.

The UniPose9D model aims to solve the challenging problem of category-agnostic 9D object pose estimation by using DINOv2 and PointNet features to predict NOCS coordinates from observed geometry. To boost accuracy, it employs a point-pair-based RANSAC N-hop Kabsch-Umeyama algorithm with an adaptive threshold.

This model is further enhanced by flow matching to manage symmetric ambiguities and uses curated public datasets to build a large training set, allowing it to perform well on novel objects and unseen categories in real-world scenarios.

Refusal-Gated Decoding introduces a method that preserves the refusal behavior of large language models even when using high-temperature sampling for greater diversity. This sequential decoding approach runs a short greedy probe that checks compatibility with learned refusal prefixes, returning the greedy response if compatible and otherwise falling back to sampling from the original high-temperature distribution.

This technique successfully raised the preservation rate of greedy refusals from ninety-one point nine percent under direct sampling to ninety-eight point three percent on average, adding only a small amount of latency to non-refusal requests.

RefGC-SR squared is a new post-processing task designed to simultaneously perform super-resolution and artifact refinement by reusing the original high-resolution reference image. It addresses the loss of fine details that occurs when downsampling high-resolution images before generation, which can lead to identity distortion and texture loss in generated content.

The training for this method involves constructing a novel triplet dataset where real-world HRRI and HRGT pairs are curated, and an LRGI is synthesized using a Diptych-Conditioned RefGC Generator. This generator uses two ControlNets, one for appearance control based on the HRRI and another for pose control based on the HRGT edge map.

The model architecture incorporates a frequency-adaptive Mixture of LoRA Experts that dynamically adjusts weights between low-frequency and high-frequency experts based on layer hierarchy, guided by a frequency-based loss function that combines terms aligning global structure with

Today's papers

The papers

Important terms

quantum-safe cryptography
This refers to new cryptographic algorithms designed to protect data against future quantum computers. The focus is on creating practical tools, like the 'quantum-safe' Python library, to implement these standards.
multimodal intelligence gap
Large language models struggle with non-textual sensory data like audio and video. They often fail to grasp cultural nuances embedded in music or sound effects, highlighting a need for deeper contextual perception.
SchemaGraphSQL
This technique uses graph algorithms to link database schema elements, helping LLMs generate much more accurate SQL queries without needing extensive training. It makes querying complex databases practical.
PhysCaP
This research focuses on teaching robotic agents to understand physical properties through active interaction rather than just imitation. It allows robots to handle novel scenarios by estimating physics directly from movement feedback.