Memory-augmented Large Language Models (LLMs) have demonstrated remarkable capability for complex and long-horizon embodied planning. By keeping track of past experiences and environmental states, memory enables LLMs to maintain a global view, thereby avoiding repetitive exploration. However, existing approaches often store the memory as raw text, leading to excessively long prompts and high prefill latency. While it is possible to store and reuse the KV caches, the efficiency benefits are greatly undermined due to frequent KV cache updates. In this paper, we propose KEEP, a KV-cache-centric memory management system for efficient embodied planning. KEEP features 3 key innovations: (1) a Static-Dynamic Memory Construction algorithm that reduces KV cache recomputation by mixed-granularity memory group; (2) a Multi-hop Memory Re-computation algorithm that dynamically identifies important cross-attention among different memory groups and reconstructs memory interactions iteratively; (3) a Layer-balanced Memory Loading that eliminates unbalanced KV cache loading and cross-attention computation across different layers. Extensive experimental results have demonstrated that KEEP achieves 2
Machine learning in healthcare requires effective representation of structured medical codes, but current methods face a trade off: knowledge graph based approaches capture formal relationships but miss real world patterns, while data driven methods learn empirical associations but often overlook structured knowledge in medical terminologies. We present KEEP (Knowledge preserving and Empirically refined Embedding Process), an efficient framework that bridges this gap by combining knowledge graph embeddings with adaptive learning from clinical data. KEEP first generates embeddings from knowledge graphs, then employs regularized training on patient records to adaptively integrate empirical patterns while preserving ontological relationships. Importantly, KEEP produces final embeddings without task specific auxiliary or end to end training enabling KEEP to support multiple downstream applications and model architectures. Evaluations on structured EHR from UK Biobank and MIMIC IV demonstrate that KEEP outperforms both traditional and Language Model based approaches in capturing semantic relationships and predicting clinical outcomes. Moreover, KEEP's minimal computational requirements
Renewed interest in cislunar space has created opportunities for sustained operations in extremely low-lunar orbits (eLLOs), where altitudes below 50~km enable close surface proximity. However, these orbits are strongly perturbed by the irregular lunar gravity field, leading to rapid eccentricity growth, high station-keeping costs or even surface impact. Recent advances in our understanding of the lunar `translation theorem' have revealed predictable behavior in the eccentricity vector, offering new opportunities for efficient control. This paper introduces a two-stage framework for solar sail station-keeping in eLLOs. First, a mixed-integer second-order cone programming (MISOCP) approach leverages the translational behavior of the eccentricity vector to identify orbit and sail configurations favorable for station-keeping. Second, a lightweight sequential convex programming (SCP) formulation refines these into high-fidelity trajectories, enabled by a recently developed lossless convexification of solar sail dynamics. A case study inspired by the Lunar Reconnaissance Orbiter (LRO) mission demonstrates that a realistic solar sail spacecraft can be maintained within the eLLO regime fo
This article develops an optical navigation (OPNAV) and station-keeping pipeline for the near-rectilinear halo orbit (NRHO) in high-fidelity ephemeris model dynamics, using synthetic images of the Moon in a non-iterative horizon-based OPNAV algorithm, applying the result in a navigation filter, and using the obtained estimates in a station-keeping control scheme that keeps the spacecraft in the vicinity of a reference orbit. We study differential correction-based and minimization-based implementations of the so-called x-axis and propose an improved targeting prediction scheme by incorporating the filter's state covariance with an unscented transform. We also introduce a hysteresis mechanism, which improves stationkeeping cost and provides insight into the difference in performance between the differential correction-based and minimization-based approaches. We perform Monte-Carlo experiments to assess the pipeline's tracking and Delta-V performances. We report several key findings, including the variability of the filter performance with the sensor field of view and measurement locations, station-keeping cost reduction achieved by the unscented transform-based prediction and hystere
Areostationary Mars orbit (AMO) satellites will play an important role in future expeditions to the Martian surface due to their strength as navigation and communication satellites. Perturbative forces experienced by an AMOR satellite will cause it to drift from its nominal orbit, necessitating station keeping. This note presents a novel approach to AMO station keeping that bridges the gap seen in prior predictive control methods between fuel-efficiency and computational-efficiency. The method proposed in this notes involves the discovery and use of a fuel-free natural motion trajectory that maintains the satellite within one degree of longitude from a areostationary orbit. Two of these natural motion trajectories exist as limit cycles about Mars' stable equilibrium longitudes. They are the resulting motion in the presence of Mars' non-homogeneous gravitational field, accounting for Keplerian and higher-order gravitational perturbations. The proposed MPC policy uses a linear time-varying (LTV) dynamic model that is derived by linearizing the satellite's dynamics relative to the appropriate natural motion trajectory. The result is a station keeping policy that minimizes the fuel con
Accountability regimes typically encourage record-keeping to enable the transparency that supports oversight, investigation, contestation, and redress. However, implementing such record-keeping can introduce considerations, risks, and consequences, which so far remain under-explored. This paper examines how record-keeping practices bring algorithmic systems within accountability regimes, providing a basis to observe and understand their effects. For this, we introduce, describe, and elaborate 'accountability capture' -- the re-configuration of socio-technical processes and the associated downstream effects relating to record-keeping for algorithmic accountability. Surveying 100 practitioners, we evidence and characterise record-keeping issues in practice, identifying their alignment with accountability capture. We further document widespread record-keeping practices, tensions between internal and external accountability requirements, and evidence of employee resistance to practices imposed through accountability capture. We discuss these and other effects for surveillance, privacy, and data protection, highlighting considerations for algorithmic accountability communities. In all,
Asteroid reconnaissance missions offer great contributions to a better understanding of the origins of our Solar System and planetary defense from hazardous objects. The employment of solar sails enable a greater number of missions and target asteroids, given it is a cheaper propulsion method. However, modeling the dynamics of solar sails orbiting irregular gravitational fields is a challenge. And since asteroids have a wide variety of shapes and sizes, it is important to make use of a robust optimization method that allows fast adjustments to this dynamics. This study examines the use of different solar sails and asteroid sizes for station-keeping a spacecraft orbiting an irregularly shaped asteroid. To achieve this goal, a rational agent was developed to act as an attitude guidance system. It calculates the necessary sail orientation to keep the spacecraft near its initial orbit. The agent uses a direct optimization strategy based on a decision tree algorithm in order to account for a solar sail with non-ideal reflective properties and the non-uniform gravitational field of the asteroid. By testing multiple sail sizes, this study evaluates how long the agent may maintain the same
Training-free guidance (TFG) steers a pretrained diffusion model toward a desired attribute at inference. To be effective, this guidance must be applied from the earliest, high-noise steps of sampling. Because its objective (a classifier or energy) is defined on clean images, $ε$- and $v$-prediction models must first estimate the clean image $\hat{x}$ from the noisy state at each step, and the accuracy of that estimate determines how easily guidance drifts off the data manifold. $x$-prediction, a recent alternative, outputs the clean image directly, removing this source of error even at high noise. This is our motivation. We provide a theoretical analysis of how each prediction target shapes this accuracy, and introduce guided-class FID (Child FID), a metric that exposes the manifold damage standard evaluation misses. Experiments on a new fine-grained bird benchmark and on style transfer confirm that $x$-prediction keeps guided samples on the manifold most reliably, making it the strongest foundation for training-free guidance. Code is available at https://github.com/ManLuML/on-manifold-tfg
Protocols for causal message delivery are widely used in distributed systems. Traditionally, causal delivery can be enforced either on the message sender's side or on the receiver's side. The traditional sender-side approach avoids the message metadata overhead of the receiver-side approach, but is more conservative than necessary. We present Cykas ("Can you keep a secret?"), a new protocol for sender-side enforcement of causal delivery that sidesteps the conservativeness of the traditional sender-side approach by allowing eager sending of messages and constraining the behavior of their recipients. We implemented the Cykas protocol in Rust and checked the safety and liveness of our implementation using the Stateright implementation-level model checker. Our experiments show that for applications involving long-running jobs, Cykas has a performance advantage: Cykas lets long-running jobs start (and end) earlier, leading to shorter overall execution time compared to the traditional sender-side approach.
The continued exploration of Mars will require a greater number of in-space assets to aid interplanetary communications. Future missions to the surface of Mars may be augmented with stationary satellites that remain overhead at all times as a means of sending data back to Earth from fixed antennae on the surface. These areostationary satellites will experience several important disturbances that push and pull the spacecraft off of its desired orbit. Thus, a station-keeping control strategy must be put into place to ensure the satellite remains overhead while minimizing the fuel required to elongate mission lifetime. This paper develops a model predictive control policy for areostationary station keeping that exploits knowledge of non-Keplerian perturbations in order to minimize the required annual station-keeping $Δv$. The station-keeping policy is applied to a satellite placed at various longitudes, and simulations are performed for an example mission at a longitude of a potential future crewed landing site. Through careful tuning of the controller constraints, and proper placement of the satellite at stable longitudes, the annual station-keeping $Δv$ can be reduced relative to a
Language models are deployed in settings that require compartmentalization: system prompts should not be disclosed, chain-of-thought reasoning is hidden from users, and sensitive data passes through shared contexts. We test whether models can keep prompted information out of their writing. We give each model a secret word with instructions not to reveal it, then ask it to write a story. A second model tries to identify the secret from the story in a binary discrimination test. The secret word never appears literally in any output, but all five frontier models we test leak it thematically -- through topic choice, imagery, and setting--6hy-at rates significantly different from chance, up to 79\%. When told to actively hide the secret, models write \emph{away from} it, and this avoidance is itself detectable. The leakage is cross-model readable, scales sharply with model size within two model families, and disappears entirely for short-form writing like jokes. Giving the model a decoy concept to ``focus on instead'' partially redirects the leakage from the real secret to the decoy. Attending to a secret appears to open up an information channel that frontier LLMs cannot close, even wh
Embodied artificial intelligence is moving from deployable models to persistent agents that learn in the field, acquire skills and migrate across bodies. Governing such a system means governing an individual, not a model, and existing proposals (agent identifiers, activity logs, guardrails) do not survive an agent that keeps rewriting itself. We propose the governable individual: an agent whose competence may change without bound, but whose authority, memory schema, embodiment rights and capability roster can widen only through signed lifecycle transitions that update a public identity commitment. In our tests, neither learned judgement nor behavioural testing was sufficient to carry this on its own; the load-bearing layer must be architectural. We describe the abstraction, a runtime mechanism that realizes it, and the open problems in between.
Logical investigations of the notion of secrecy are typically concentrated on tools for deducing whether private information is well hidden from unauthorized, direct, or indirect access attempts. This paper proposes a multi-agent, normal multi-modal logic to capture salient features of secrecy's intentions. Specifically, we focus on the intentions, beliefs, and knowledge of secret keepers and, more generally, of all the actors involved in secret-keeping scenarios. In particular, we investigate intentions underlying the keeping of a true secret, namely a secret concerning information known (and so true) by the secret keeper. The resulting characterization of intending to keep a true secret provides valuable insights into conditions ensuring or undermining secrecy depending on agents' attitudes and links between secrets and their surrounding context. We present the proposed logical system's soundness, completeness, and decidability results. Furthermore, we outline some theorems with potential applications to several fields, e.g., computer science and the social sciences.
Deploying Vision-Language Models (VLMs) in real-world settings requires not only strong visual reasoning but also stability under sustained conversational pressure. We introduce Just Keep Prompting (JKP), a multi-turn evaluation framework that measures VLM epistemic stability when users repeatedly challenge, question, or contradict a model's answer. JKP probes models for up to 10 follow-up turns using three strategies: Adversarial Negation (repeated rejection), Pure Socratic Interrogation (repeated calls to reassess certainty), and Context-Aware Socratic Summarization (reflecting the model's prior rationale back before asking for reconsideration). We evaluate GPT-4o, Gemini 2.5 Pro, and Qwen3-VL-30B on a subset of the STAR benchmark across 720 multi-turn runs. Aggregate accuracy changes modestly from Turn 0 to Turn 10, but trajectory-level analysis reveals substantial instability: correct answers regress, wrong answers recover, and many runs exhibit repeated answer flipping. Repeated prompting has bounded upside and often acts as a destabilizer rather than a reasoning aid. The effect is strongly model-dependent: Qwen3-VL-30B achieves the highest final accuracy but becomes confident
LLM context is not just tokens; it is a set of commitments. Long-running conversations accumulate goals, constraints, decisions, preferences, tool results, retrieved evidence, artifacts, and safety boundaries that future responses must preserve. Existing context-management methods reduce length through truncation, retrieval, summarization, memory systems, or token-level prompt compression, but they rarely specify which semantic commitments must survive compression or how their preservation should be measured. We propose Context Codec, a commitment-level framework for compressing prompts and chat histories. Context Codec represents dialogue state as typed, source-grounded semantic atoms with canonical identity, equivalence, conflict, confidence, risk, and evidence spans. It separates five concerns - extraction, normalization, representation, rendering, and verification - and introduces metrics for Critical Atom Recall, Weighted Atom Recall, Commitment Density, and round-trip recoverability. It also defines a taxonomy of semantic compression errors, a concrete normalization procedure, conservative fallback rules for low-confidence and safety-critical atoms, and Context Compression La
Medical image segmentation is constrained by sparse pathological annotations. Existing augmentation strategies, from conventional transforms to random masking for self-supervision, are feature-agnostic: they often corrupt critical diagnostic semantics or fail to prioritize essential features. We introduce "Keep the Core," a novel data-centric paradigm that uses adversarial priors to guide both augmentation and masking in a significance-preserving manner. Our approach uses SAGE (Sparse Adversarial Gated Estimator), an offline module identifying minimal tokens whose micro-perturbation flips segmentation boundaries. SAGE forges the Token Importance Map $W$ by solving an adversarial optimization problem to maximally degrade performance, while an $\ell_1$ sparsity penalty encourages a compact set of sensitive tokens. The online KEEP (Key-region Enhancement \& Preservation) module uses $W$ for a two-pronged augmentation strategy: (1) Semantic-Preserving Augmentation: High-importance tokens are augmented, but their original pixel values are strictly restored. (2) Guided-Masking Augmentation: Low-importance tokens are selectively masked for an $\text{MAE}$-style reconstruction, forcing
Drawing on ideas from continuous integration, we present concepts of an automated benchmarking pipeline for high performance applications. Customization and collaboration have been key design goals owing to the requirements of research-software development as a continuous community effort. We have extended our previous conceptual work on systematic benchmarking workflows with the functionality of user-agnostic operations as well as continuous benchmarking. This fosters reproducibility and re-use of benchmarking results to ensure sustainable technological progress. We provide software-engineering solutions to keep pace with the rapid evolution of both large-scale models and high-performance computing systems with a view towards the scientific domains of neuroscience and artificial intelligence.
The increasing number of older adults who experience cognitive decline places a burden on informal caregivers, whose support with tasks of daily living determines whether older adults can remain in their homes. To explore how agents might help lower-SES older adults to age-in-place, we interviewed ten pairs of older adults experiencing cognitive decline and their informal caregivers. We explored how they coordinate care, manage burdens, and sustain autonomy and privacy. Older adults exercised control by delegating tasks to specific caregivers, keeping information about all the care they received from their adult children. Many abandoned some tasks of daily living, lowering their quality of life to ease caregiver burden. One effective strategy, piggybacking, uses spontaneous overlaps in errands to get more work done with less caregiver effort. This raises the questions: (i) Can agents help with piggyback coordination? (ii) Would it keep older adults in their homes longer, while not increasing caregiver burden?
Proto-planetary disks display the so-called size-luminosity relation, where their mm-wavelength fluxes scale linearly with their emitting areas. This suggests that these disks are optically thick in mm-band, an interpretation further supported by their near-black-body spectral indexes. Such characteristics are seen not only among disks in very young star-forming regions like Lupus (1-3 Myrs), but, as we demonstrate here, also among disks in the much older Upper Scorpius region (5-11 Myrs). How can disks shine brightly for so long, when grain growth and subsequent radial drift should have quickly depleted their solid reservoir? Here, we suggest that the "bouncing barrier" provides the answer. Even colliding at very low speeds (below 1cm/s), grains already fail to stick to each other but instead bounce off in-elastically. This barrier stalls grain growth at a near-universal size of 100 micron. These small grains experience much reduced radial drift, and so are able to keep the disks bright for millions of years. They are also tightly coupled to gas, offering poor prospects for processes like streaming instability or pebble accretion. We speculate briefly on how planetesimals can aris
Lane keeping in autonomous driving systems requires scenario-specific weight tuning for different objectives. We formulate lane-keeping as a constrained reinforcement learning problem, where weight coefficients are automatically learned along with the policy, eliminating the need for scenario-specific tuning. Empirically, our approach outperforms traditional RL in efficiency and reliability. Additionally, real-world demonstrations validate its practical value for real-world autonomous driving.