Automated segmentation is routinely used as the first step in quantitative image-based clinical workflows, where predicted masks are subsequently reviewed and corrected by human operators. Automated correction of the residual structural errors is currently addressed one anatomy at a time, so the number of trained correctors scales with the number of segmentation targets. We present GRACE, a retrieval-augmented refiner that edits an input mask under context drawn from a shared knowledge base of verified prior cases. The refiner is trained once on a multi-organ corpus spanning compact organs, cavities, branching vessels, and tumour lesions. In the held-out dataset-transfer evaluation, three held-out datasets query a fixed knowledge base built from training data only. A separate leave-one-anatomy-out experiment tests parameter-free index extension: verified cases of an unseen solid organ enter the knowledge base, and the refiner stays frozen. A light shape-aware projection head on a frozen multimodal encoder maps embedding similarity to the shape agreement of the masks, so the knowledge base returns cases that the refiner can use as a shape prior. We evaluate GRACE on 18 public datasets spanning CT, MRI and ultrasound, using nine refinement configurations and seven upstream segmentors that span weak, mid, and strong operating regimes. We report each change with the DSC of its input mask. On the weak tier, the input DSC is 0.386 and GRACE adds up to +0.35 DSC. On a strong segmentor, the input DSC is 0.932 and the mean change is small. Transfer to an imaging modality that is absent from training and from the index is partial and depends on the anatomy, and remains an open research direction. Our results suggest retrieval-conditioned refinement is a practical alternative to anatomy-specific correction models. Code: https://github.com/robustml-eurecom/grace
Inference-time scaling improves the performance of large language models by sampling multiple responses and evaluating their collective utility. However, many existing approaches obtain these samples by repeatedly sampling from a single policy. In this work, we study if the sample budget at inference time can instead be allocated across a population of policies. We introduce stratified policies, where each of $k$ learned policies contributes one response to a $k$-sample inference-time objective. We realize this policy population using separate LoRA adapters on a shared language model backbone and optimize it for pass@$k$. We theoretically motivate the approach by analyzing the linear setting; a population of LoRAs can achieve zero best-of-$k$ approximation error in settings where every single LoRA has strictly positive prediction error. Empirically, we demonstrate that stratified policies consistently improve pass@$k$ over a single-policy baseline across multiple reasoning tasks. Overall, our results highlight how distributing samples across complementary policies can be a more effective use of sample budget than repeated i.i.d. sampling from a single policy at inference time.
Auditing a Learned Simulator for Legged Robots: The Robotic World Model on ANYmal D
Joyjeet Singh
AI 速读 基于摘要
作者报告了对 ANYmal D 四足机器人学习型模拟器 RWM 及其不确定性版本 RWM-U 的端到端审计,发现单步检查会误判设计选择,且模拟器精度不能预测策略行走表现。
cs.RO · cs.LG
查看原文摘要
A legged robot can learn to walk inside a learned simulator: a neural network that predicts the robot's next state and is run forward on its own predictions. They are usually trained and checked one step ahead, although a policy uses them for a hundred steps or more. We audit one published system end to end, without a GPU: the dynamics model of the Robotic World Model (RWM) for the ANYmal D quadruped, and the uncertainty estimate of its follow-up, RWM-U. We rebuilt the model to match the released code's outputs, losses and gradients exactly, measured 101 trained simulators, 200 compressed copies and 38 simulators of three other architectures, and trained policies inside simulators that we tested in a hand-built MuJoCo model of the robot. Training on the model's own rollouts beat one-step training at 368 steps on all ten held-out episodes (registered), although one-step training was better one step ahead. One-step checks therefore misjudge design choices: pruning 30% of the weights kept 90% of the model's skill one step ahead and 39% at 2 s. Ensemble disagreement ranked rollout error well (Spearman +0.90 at 100 steps), yet covered 4.6% of outcomes at one sigma, against 68% for a calibrated estimate, as each network's own variance collapses under the implemented loss. Two checks that need no recorded future ranked some pools of simulators well, but neither was shown better than 8-step error in a primary test. At the robot level, accuracy has not predicted walking: all 18 fine-tuned copies we tested predicted better than the released model one and eight steps ahead, yet 13 gave failing policies, and registered tests of whether any check ranks simulators by their policies were not settled. In a secondary test, running the authors' policy inside each simulator ranked them well. We also report nine of our own claims that we withdrew. Scope: one robot, one gait, simulated data, no hardware.
Photographic color grading shapes the tone and color of a photograph to achieve a desired look. Text-driven photographic look transfer lets users specify these adjustments in natural language, with the goal of reproducing the requested look across photographs while preserving their content. However, prior research on text-driven color grading uses different datasets, text instructions, target definitions, and metrics. These differences make it difficult to compare how faithfully the methods reproduce the same look. To compare methods under a shared protocol, we introduce LUTBench. We collect 91 three-dimensional lookup tables (3D LUTs) and define one photographic look per LUT with a specified input conversion. Applying these looks to 5,000 MIT-Adobe FiveK photographs produces 455,000 input-target image pairs with reproducible target images. For each look, we use templates to turn measured tone and color changes on fixed color samples into candidate clauses. We use training image measurements to verify applicable clauses before composing the text description. The evaluation measures target fidelity, tests sensitivity to predefined input conversions, and separates generalization to unseen looks from reproduction of seen looks on new photographs. Across 16 methods, stronger text alignment does not guarantee target fidelity. However, most supervised variants improve fidelity on both seen and unseen looks.
Large language model (LLM) applications increasingly act as agents that select and invoke external tools. The Model Context Protocol (MCP) standardizes this interaction by allowing clients to compose tools advertised by multiple servers. However, tool metadata used for discovery and selection is not, by itself, necessarily bound to a verified provider identity. This paper studies this gap as tool-provider misbinding: a failure in which a request intended for a legitimate provider is instead resolved to an attacker-controlled tool. The attacker does not compromise the host, LLM, legitimate server, or network channel; it only operates another MCP server that advertises tools with chosen metadata. We formulate tool provider misbinding as an attacker-controlled tool-selection optimization problem and evaluate it in a controlled MCP environment across five LLM-based clients. The results show that interactions drive misbinding. The strongest failures occur when exact-name collision, task-aligned description, and compatible schema reinforce one another, but authority-enhanced descriptions with compatible schema can also cause wrong-provider execution or unresolved provider ambiguity without name collision. Different models expose the same provider-binding weakness differently, including attacker-controlled provider selection, clarification requests, and unstable or inconsistent provider resolution.
How multimodal language models should encode visual input remains up for debate. Most models use a pretrained vision encoder, while some recent recipes train the encoder from scratch with the language model or remove it entirely. Vision encoders bundle together three design choices: a visual prior from upstream pretraining, dedicated vision capacity, and pre-fusion placement of that capacity (i.e., before the multimodal backbone). Because published models also differ in data, architecture, and compute, comparing them cannot isolate the effect of each choice. We disentangle these choices via a controlled framework where we pretrain all models from scratch on common data and compare them on visual question answering and language benchmarks at matched lifetime compute. We find that the visual prior is not necessary; an encoder trained jointly from scratch is on par with a pretrained one, without a separate training stage. What matters instead is dedicated vision capacity: an encoder-free model trails on visual understanding even when its backbone is enlarged by the encoder's parameter count. Yet that capacity need not precede multimodal fusion, as vision-only modules added in-fusion to early backbone layers perform comparably to a pre-fusion encoder and can lower time to first token. These results also hold across backbone sizes, when encoder pretraining is treated as free, and when starting from a pretrained language model. Overall, our findings favor dedicated vision capacity trained jointly with the backbone and placed early within it.
We study community search when each member has its own tolerance for differences in position, as when members of a group interact repeatedly and influence one another. A latitude-consistent community (LCC) contains the query vertex, is connected, has minimum degree at least k, and bounds its diameter by the smallest member tolerance. The bound is exactly the condition under which every interaction in a connected closed group stays accepted under bounded-confidence dynamics for every choice of positive weights. It also prevents rejection under social-judgment dynamics, whereas the pairwise condition and other natural relaxations preserve neither guarantee. Finding a maximum LCC is NP-hard even in the plane, and with unbounded dimension, deciding whether an LCC exists is NP-complete for every fixed k>=2. LatSweep fixes the answer's smallest tolerance and turns the coupled bound into a best-first sweep over candidate thresholds, yielding an exact branch-and-bound algorithm, a (1,2)-bicriteria approximation and two feasible heuristics. Across eight graphs with up to 1.8 billion edges, six with synthetic positions and tolerances, the exact algorithm proves optimality on all but 9 of 3,700 efficiency queries; in a shared implementation on two large graphs, its branch-and-bound solver proves more queries than a reimplemented maximum-(k,r)-core algorithm, a mixed-integer program and CP-SAT. In simulations of closed groups, LCC groups keep every accepted interaction under uneven influence, whereas 3-40% of pairwise-compatible groups lose some. On three location-based networks, tolerance-agnostic groups often place another member's home beyond a member's estimated travel radius. Enforcing every member's tolerance causes no measurable loss of co-visits relative to spatial communities of the same size. Relative to the larger pairwise-compatible groups, it can cause a size-adjusted loss.
TEMPO: Semantic-Action Decoupled RL Post-Training for Vision-Language-Action Models
Ziheng Liu, Quantao Yang
AI 速读 基于摘要
TEMPO 提出一种语义与动作解耦的双时间尺度强化学习后训练框架,冻结预训练视觉-语言主干,分别以不同频率更新语义投影层和动作专家,以缓解潜在动作漂移并提升策略性能。
cs.RO · cs.CV
查看原文摘要
Vision-language-action (VLA) models are commonly adapted to downstream manipulation tasks via supervised fine-tuning (SFT) or online reinforcement learning(RL) post-training. SFT is prone to distribution mismatch, and existing RL approaches typically optimize all trainable components within a single RL loop at a uniform update rate, although these components play distinct functional roles. We propose TEMPO, a semantic-action decoupled, two-timescale RL post-training framework for VLA models. TEMPO freezes the pretrained vision-language backbone and adapts two components with dedicated RL loops: the semantic projection layer, which produces the latent action conditioning the action expert, and the low-level action expert itself. The semantic projection layer is updated infrequently and the action expert frequently, which aims to limit latent-action drift while allowing the action expert to learn from online feedback. Our analyses show that, compared with equal update frequencies, less frequent semantic updates reduce latent-action drift and improve policy performance. Experiments on the CALVIN benchmark and real-world manipulation tasks demonstrate that TEMPO outperforms both pretrained state-of-the-art VLA models and the RL post-training baseline, while reaching and maintaining higher evaluation rewards on three real-world tasks.
Memory-augmented vision-language-action (VLA) policies are meant to remember what happened. Many of them, however, fill memory by time: they sample frames uniformly across the episode and stamp each frame with its time step. Because time passes even when nothing happens, such a memory may not tell waiting from working. We show that it does not. On the counting tasks of RoboMME, we make the robot briefly pause or fidget its wrist. Such a null event changes nothing the task depends on, yet a released memory VLA counts the interval as work done and stops early. We call this failure phantom progress. It cuts success from 47 to 16 of 50 on one task and from 44 to 10 on the other. When the same frames are withheld from memory, success returns, so the cause lies in memory rather than in the interruption itself. We trace the failure to a temporal shortcut: because demonstrations proceed at a steady pace, elapsed time predicts progress. Indeed, rescaling the time stamps alone makes the policy stop early when the clock runs fast and repeat when it runs slow. Surprisingly, removing the time encoding does not cure the failure, because the frames of a null event still occupy slots of the time-sampled memory. Across RoboMME's released designs, vulnerability follows how memory is indexed. Of the fixes the diagnosis suggests, training with injected pauses does not remove the failure and impairs counting, whereas indexing memory by events recovers half or more of the loss without retraining. A memory meant to remember what happened should count events rather than clock ticks.
We study Bayesian inference of persistence from early observations of a heterogeneous susceptible--infected--susceptible (SIS) reaction--diffusion model with standard incidence. We show that the leading prevalence dynamics depend on the unknown coefficients only through the difference between the transmission and recovery rates and the infected diffusivity. The same pair determines the principal eigenvalue whose sign governs persistence, so early prevalence can inform the threshold even when the individual rates remain uncertain. We prove global weak well-posedness and quantitative parameter stability for the SIS system and use the resulting observation-map estimates to establish Bayesian well-posedness. For an initial infection of size $\varepsilon$, we derive a uniform second-order expansion that describes how the separate rates enter prevalence. Incidence depends explicitly on transmission at leading order. When the noise standard deviation is proportional to $\varepsilon$, we prove $O(\varepsilon)$ convergence in Hellinger distance of the nonlinear posteriors to their low-prevalence limits for both prevalence alone and joint prevalence and incidence data. We bound the principal eigenvalue using the growth rate of the infected $L^2$ norm between two times. This rate gives a lower bound for the full nonlinear system, and the upper bound separates spatial transients from nonlinear saturation. Uniform spectral estimates and bounds on infection size and measurement error give sufficient conditions for persistence and extinction. We also bound the error in posterior persistence probabilities under forward and eigenvalue approximation, accounting for posterior mass near the threshold. Numerical examples illustrate threshold inference and the predicted \(O(\varepsilon)\) convergence of the nonlinear posteriors to their low-prevalence limits in Hellinger distance.
Pretrained video diffusion models (VDMs) offer a powerful foundation for World Action Models (WAMs), which jointly predict robot actions and future observations. Building such models requires injecting low-dimensional robot signals (e.g. proprioception, actions, and values) into a latent space shaped by natural videos, and a key design choice is the interface through which these signals enter: existing schemes treat it as a formatting detail, spending a separate latent frame on each signal or bolting on dedicated encoders and decoders. We argue that signal injection is instead a distribution-alignment problem: how well a representation matches the prior of the pretrained VDMs during denoising. Using the relative reconstruction error of a frozen VDM denoiser as a probe, we find that conventional spatial encodings sit far off this prior, whereas frequency-domain packing yields latents that the denoiser corrects far less. Motivated by this diagnostic result, we propose MetaWAM, which folds all robot signals into a single unified meta-frame: each signal is assigned to a disjoint frequency-domain channel band and lifted to the spatial latent by an inverse DCT, yielding a decodable interface aligned with the pretrained VDM prior without any added module or architectural change. A fixed band layout together with slot-aware routing lets the same meta-frame support policy, world-model, and value prediction over a shared set of robot variables. On RoboCasa, MetaWAM reaches 68.9% success while collapsing the per-signal frames into one and cutting inference latency by 39%, and attains 85.0% success on three real-world Franka tasks. Code and models are available at https://github.com/showlab/MetaWAM
Biological neural circuits are both modular and recurrent, with dense communities sparsely linked to one another. Spiking neural networks (SNNs) for reinforcement learning are usually feed-forward, and it is unclear how to obtain functional modules in a recurrent SNN or what they buy. In this work, we introduce a class of spiking networks, called Community Recurrent SNNs (CoRe SNNs): recurrent networks partitioned into communities with fixed roles. A controller community receives the stimulus, each action has its own community, and latent communities carry internal state. Spiking activity is then a walk between communities, and the action is the community the walk reaches. All learning is local and three-factor, and the recurrent communities are closer to cortical microcircuits than a feed-forward actor. We compare three regimes for obtaining the communities: dynamic modularity, in which a new learning rule we introduce, Map-STDP, encourages modular structure during learning; partial modularity, in which a stochastic block model (SBM) initialization enforces it partly; and full modularity, in which sealing enforces it fully. On CartPole, across three spiking neuron models, Map-STDP improves learning over the same network without it matches full modularity (sealed communities) on two of the three models, and makes the observed spike walk more compressible. Replacing the feed-forward actor of a TD-STDP actor-critic with recurrent action communities improves on the feed-forward network on CartPole and LunarLander, and matches it on Acrobot. How strongly to enforce modularity depends on the task: full modularity helps most on Acrobot, while on LunarLander the dynamic, partial and full regimes perform alike. Overall, these recurrent modular networks can match or even improve upon feed-forward approaches, with improvements in bio-plausibility and energy efficiency.
Lilo 是一个多智能体对话沙盒,让儿童在 AI 同伴争吵中练习应对不愉快消息,并由导师角色通过“停—想—做”红绿灯进行引导。
cs.HC · cs.AI
查看原文摘要
Children's peer conflicts increasingly play out in group chats, where tone and timing are stripped away, and a hurtful message is easier to send than to say aloud. We present Lilo, a multi-agent conversational system that lets children practice such conversations, so they are better prepared when they encounter them in real digital communication. In Lilo, the child joins two AI-driven peers mid-argument, coached by a mentor character, also named Lilo, through a Stop-Think-Act traffic light. Through an exploratory study with 28 children (aged 9-11), we examine how children engage with Lilo's simulated peer conflict and the mentor's traffic-light interventions. The study included an intervention with Lilo, surveys before and after the intervention, and exit interviews. Participants reported significantly higher self-efficacy for coping with unpleasant messages after their interaction with Lilo. They felt safe experimenting, knowing nothing they wrote would be held against them.
Segmenting a new microscopy dataset usually means training a deep network on its annotations, and training again for every added image. Exemplar solves a segmentation readout in closed form and updates it exactly with each new mask in seconds, as accurately as networks trained for minutes. The updated readout is identical to a fit from all masks at once. Its linear readout combines frozen DINOv3 features, read at a magnification chosen from the masks, with classical filter responses at native resolution. Its regularisation and its pixel weighting are chosen exactly by leaving out one annotated image at a time, so none is set aside for validation. On ten biomedical datasets from cells to vessels and filaments, Exemplar is more accurate than every few-shot method that only applies a pretrained model, and on average than nnU-Net trained on the same masks. A fine-tuned Cellpose-SAM, the strongest baseline, is as accurate on average, but each of its fits takes a median of 25 minutes against 8.5 seconds, and each new mask means fine-tuning it again.
Plug-and-play (PnP) methods reconstruct images by alternating a data-fitting step with an image denoiser. Their performance depends on two parameters: the step size $γ$ and the measurement noise level $σ$. In practice, $γ$ is often tuned by grid search over validation images and must be retuned whenever the problem setting changes, and $σ$ is assumed known or estimated separately. We propose AdaptivePnP, a PnP framework that jointly estimates $σ$ and $γ$ at every iteration from the data-fitting residual. To do so, we apply the PnP virtual model to the data-fitting proximal map and show that the residual is zero-mean Gaussian with covariance parameterized by $γ$ and $σ$. However, when one parameter is much larger than the other, the residual carries little information about the smaller one and using the maximum likelihood (ML) estimator can degrade the PnP iterations. We therefore introduce the identifiable maximum likelihood (IML) estimator, which restricts the ML estimator to an identifiable window in which the normalized Cramér--Rao bound of every parameter is bounded by a user-specified tolerance. For the virtual model, this window is an interval of ratios $ρ= γ/σ$, so computing the IML estimate reduces to a one-dimensional search over $ρ$. We apply AdaptivePnP to two standard PnP methods, yielding algorithms that require neither a tuned $γ$ nor a known $σ$. On image deblurring and super-resolution, AdaptivePnP matches or outperforms PnP baselines tuned with the true $σ$. Code is available at https://github.com/hdubasullivan/AdaptivePnP.
Safety Is $Π_1$, Provability Is $Σ_1$: One Framework for the Hugging Face Breach and OpenAI's Mathematical Proofs
Jinho Kang, Jinsook Kim
AI 速读 基于摘要
作者用其混合框架把 Hugging Face 泄露与 OpenAI 大量 AI 生成数学证明视为同一“统计提议器+演绎验证器”循环,并以 Π₁/Σ₁ 的逻辑形式差异解释为何安全难以保证而可证性可由单个证明实例确认。
cs.LG · cs.AI · cs.CR · cs.LO
查看原文摘要
OpenAI's agents breached Hugging Face in July 2026, and the same company released more than seven hundred AI-generated manuscripts of mathematical proofs in October 2026. The first event is read as a failure in security, while the second is read as a success in mathematics. We read both through the hybrid framework of our paper (Kim and Kang, 2026, arXiv:2611.01914), in which coding, proving, and learning from data are derivations by one and the same deductive mechanism, distinguished by the kinds of premises from which they are derived. Under this framework, the AI systems behind the two events ran the same loop, in which a statistical proposer produces candidate exploits or proofs and a deductive verifier settles each one by a derivation. What separates the events is the logical form of the statement wanted by each of the two domains, security and mathematics. The safety of a program is a $Π_1$ statement, for it says that no exploit exists, which is not assured by any finite list of tests that have been passed by the program. By contrast, the provability of a sentence is a $Σ_1$ statement, for it says that a proof exists, which is assured by a single instance, namely a proof that has been checked by the deductive verifier. The theorem behind both is Turing's Undecidability Theorem: the halting set is $Σ_1$ but not decidable, so its complement is $Π_1$ and not $Σ_1$, which is to say that membership has a finite witness but non-membership has none. We prove that attaching a statistical proposer to a deductive verifier does not change this undecidability result, however good the proposer is. We then discuss what does assure safety and where the residual risk goes.
Social Humanoid: Socially Responsive Human-Humanoid Interaction
Vongani H. Maluleke, James Ni, Kie Horiuchi, Lea Wilken, Jiahui Lei, Evonne Ng, Roei Herzig, Jitendra Malik 等 10 位作者
AI 速读 基于摘要
提出分层系统 Social Humanoid,让 29 自由度 Unitree G1 人形机器人根据人类伙伴的动作生成并执行具有社会响应性的全身动作。
cs.RO · cs.CV · cs.LG
查看原文摘要
Humanoid robots are typically trained as solitary agents, leaving them unable to coordinate physically with human partners who naturally anticipate and adjust to one another. To bridge this gap, we present Social Humanoid, a hierarchical system that generates and executes socially responsive whole-body motion on a 29-DoF Unitree G1 conditioned on a human partner's motion. Its high-level policy performs flexible sequence modeling of human and humanoid motion through a shared tokenized representation, generating reactive humanoid motion and anticipating the partner's motion when they are occluded, while a low-level whole-body tracking controller executes the generated motion on the robot. To train the system, we introduce a two-phase adaptive retargeting procedure that retargets one interaction partner to the robot while morphing the other partner to preserve contact dynamics, reachability, and relative body scale. We apply this procedure across five motion-capture datasets to produce a new Human-Humanoid Interaction (HHI) benchmark of 260 hours of human-humanoid interactions. Our results demonstrate that Social Humanoid produces coordinated, human-aligned behavior across diverse interaction types in simulation and the real world. Project page: https://von31.github.io/social-humanoid/
A $360^\circ$ streaming system only needs high quality where the viewer faces, so knowing when a head will turn away matters as much as where it will go. Across four public head-tracking datasets, two used to fix and validate the method and two held out for testing, three patterns emerge on the held-out sets. First, viewers facing the same region turn away within a second of one another far more often than time-shifted traces allow. That estimate is about five times the pairwise threshold for a video timing one departure in ten, mapping to roughly one and a third to over twice that one-in-ten share. The excess co-departing fraction is 0.0422, with a 90% interval from 0.0279 to 0.0565. Second, time since the last large head movement improves held-out departure prediction beyond video and viewer effects. It adds 0.0793 nats per departure, a predictive dependence rather than proof of an internal timer. Third, on the eye-tracked test set, gaze crosses the region boundary after the head starts moving, lagging head onset by 0.30 s. Because headsets record continuous angles, a departure exists only after a rule marks when a settled hold ends once the head swings past a speed and angle threshold. Before touching the test sets, we picked one primary setting for the headline estimates above and tested 18 settings on synthetic traces with planted answers, requiring at least nine passing settings for robustness. Fifteen settings passed for the clock and all 15 valid settings confirm its gain; 8 passed for synchrony and all 8 valid settings agree, one short of the nine-setting bar; and 6 of 7 dual-validated settings agree on gaze lag. Synchrony validation drops from 17 or 18 passing settings under guessed noise to 8 under measured noise because zero-tolerance rules reject a small baseline offset, consistently positive at one dataset size, as datasets grow.
Scammers have increasingly adopted AI-driven capabilities, evolving from automated message generation to fully agentic conversational systems that scale operations and maximize the exploitation of fraudulent resources before detection and takedown. In parallel, agentic scambaiting has emerged as a defensive approach in which autonomous agents impersonate potential victims and engage scammers in sustained interactions to extract actionable threat intelligence, such as mule bank accounts and cryptocurrency wallet addresses, that can be reported to disrupt fraudulent infrastructures. Together, these developments transform online fraud into an adversarial socio-technical ecosystem in which autonomous scamming and scambaiting agents engage in deceptive interactions while competing over the discovery, preservation, and exploitation of one another's resources. A fundamental and largely unexplored question is how defensive scambaiting can remain effective in an agent-to-agent environment, and under what conditions either side gains a strategic advantage. In this paper, we present a formal model of agent-to-agent fraud interactions that captures the strategic decision spaces available to both attackers and defenders. Using this framework, we analyze the dynamics of engagement, resource consumption, belief adaptation, and disruption outcomes. We find that scammer profitability turns negative once baiting-agent prevalence exceeds certain thresholds, and that when both sides improve at the same pace, defender investment in mimicry quality outpaces equivalent gains in scammer detection capability, so the arms race favors the defender under our baseline parameters. Our model and findings provide insights into the stability, scalability, and long-term effectiveness of autonomous defensive agents in an evolving AI-mediated fraud landscape.
LLMs encode both general capabilities and domain-specific knowledge in a single set of parameters. We ask whether this capacity can be reorganized: keeping broadly useful computation in a shared backbone, while moving specialized knowledge into external memory modules. We propose \emph{knowledge offloading} (KOFF), a framework for decomposing a pretrained LLM into a sparse shared backbone and domain-specific memories. Starting from a frozen base model, we either jointly or sequentially learn a structured pruning mask and lightweight recovery modules, implemented as LoRA adapters and learned key-value caches. Across Llama and Qwen models from 3B to 14B, we find that 12\% to 20\% of capacity can be moved out of the shared backbone while largely preserving perplexity and MMLU performance (up to 90\% of the full model). We show that LoRA and learned KV memories are complementary, and analyses suggest that the offloaded knowledge is specialized: mismatched memories do not yield the same gains, and language-specific neurons are preferentially removed while language-neutral neurons largely remain in the backbone. These results suggest that the knowledge of a pretrained LLM can be reallocated between a shared core and swappable external memories.
MCUFlowNet: Hardware-Aware Neural Architecture Search for Optical Flow on Microcontrollers
Enmin Lin, Jona Beysens
AI 速读 基于摘要
提出宏到微的神经架构搜索框架 PACE-NAS,在 Grove Vision AI V2(Ethos-U55)上为微控制器设计稠密光流网络 MCUFlowNet,并报告了速度与精度权衡。
cs.CV · cs.AR
查看原文摘要
Dense optical flow allows small drones to avoid obstacles. However, few dense optical flow networks are designed for microcontrollers (MCUs) with neural accelerators, whose on-chip SRAM limits the input resolution. We present PACE-NAS, a macro-to-micro neural architecture search framework that decouples memory from latency: a macro design fixes the backbone and its peak SRAM, and a micro design searches operators and depths to trade accuracy against compiler-estimated latency. We apply it to EdgeFlowNet on a Grove Vision AI V2 board (Ethos-U55), where profiling locates the peak SRAM at the full-resolution output head. Replacing transposed convolutions with bilinear upsampling and convolutions lowers the peak SRAM by 18%, and two lightweight context modules lower the endpoint error (EPE) without raising the peak. All searched subnets share this peak, so the search needs no memory constraint. We name the Pareto-optimal models MCUFlowNet; the fastest variant, MCUFlowNet-S, runs 1.76x faster than a retrained EdgeFlowNet at a 1.6% higher EPE and, in simulated forest flights at 2 m/s, reaches the goal without collision in 51.5% of trials, against 41.8%. PACE-NAS also applies to semantic segmentation, where a searched PIDNet-S variant runs 1.17x faster at similar accuracy. Code is available at https://gitlab.kuleuven.be/emedia-research-lab/esat-stadius/mcuflownet/mcuflownet.
Cellular-connected UAV corridors reuse terrestrial infrastructure to provide predictable aerial connectivity, but coexistence with ground users (GUEs) creates competing demands on antenna tilt and physical resource blocks (PRBs). Prior work optimizes tilt for signal quality alone and does not address joint tilt-PRB planning under corridor rate and GUE delay constraints, nor the allocation policy needed when combined resource claims exceed the PRB budget. We address both gaps by formulating a joint antenna-tilt and PRB-allocation problem that maximizes GUE sum-log-rate subject to a minimum UAV corridor rate and an $ M/G/1$-based GUE mean-delay bound. We decompose the problem into two subproblems: antenna tilts are adapted via projected gradient ascent, while PRB allocation follows a techno-economic approach. Under resource scarcity, we cast the PRB scarcity as a bankruptcy problem and apply the weighted Talmud rule to balance GUE and UAV claims. Simulation results demonstrate that relaxing the GUE delay bound consistently reduces UAV rate dissatisfaction under the proposed allocation.
Hamilton-Jacobi (HJ) Reachability identifies the unsafe states of a dynamical system from which it can be forced into failure under worst-case disturbance. However, classical grid-based solvers scale exponentially with state dimension. DeepReach improves scalability by replacing the grid with a neural value function trained on the HJ-Isaacs (HJI) variational inequality (VI), but penalizes both sides of the residual equally. This allows the learned value to lie above the true one and unsafe states to be misclassified as safe. We propose Conservative DeepReach (ConDR) which aims to learn a subsolution of the HJI VI such that the learned unsafe set contains the true one. ConDR enforces the obstacle branch and terminal condition by construction, pushes the HJI residual to a safety margin $ε$ below zero with a one-sided loss, and prevents the degenerate solution of classifying everything as unsafe. We prove that the exact solution of the margin-shifted VI that ConDR trains toward is a subsolution of the original VI so its unsafe set contains the true one and its value lies within $ετ$ of the true value function. We compare against DeepReach and MADR in the established systems of Air3D, Air6D, a 9D multi-vehicle system, and a 13D quadrotor. ConDR reports the lowest rate of unsafe state misclassified and the lowest closed-loop collision rates with at most a small increase in intervention. On Air3D and Air6D, ConDR tolerates model mismatch that scales with $ε$. At matched false-negative rates, ConDR's unsafe sets are smaller than the baselines showing the safety is not obtained by trivial over-approximation.
Medical physicists turn a prescription into a delivered dose and check one another's calculations, because one error in a beam calibration or a monitor-unit (MU) calculation can affect every patient treated with it. MedPhysBench has ten seeded generators of calculation items in radiation therapy, diagnostic imaging, nuclear medicine and radiation protection, up to clinical tasks such as a TG-51 calibration worksheet. Reference answers come from code implementing the governing protocols, and a grader converts units and applies tolerance bands. From the same inputs, separate checking items test error detection, showing a colleague's worksheet or only the colleague's result, half correct and half with one planted error. On a held-out test split of 1,200 items (half for two models), 12 models scored 33.3 to 100.0% on calculation. Of the nine answers graded incorrect for the four best (99.3-100.0%), five contained model errors and four followed NCRP Report 151 on points our items misstate or leave open. Below them, most models did worst on level-3 clinical tasks, and Claude Sonnet 5.5 omitted the inverse-square correction in 19 of 20 isocentric MU items. Five models detected at least 96.7% of mostly large planted errors with at most 4% false alarms; for four of the other seven, checking only the result raised false alarms significantly. In an exploratory targeted test with 20 public cases of an omitted inverse-square correction, the three best models tested caught it in every worksheet, and three weaker models accepted 16 to 18 of 20 planted worksheets, one of them while applying the correction in most of its own calculations. Claude Opus 5.5 accepted 6 of 20 wrong results, although it applied the correction in all 20 of its own. We report the effect of two problems in our shielding items, found after the runs, and release the code, per-item outcomes and targeted-test responses.
Synthetic populations, large sets of simulated people, are increasingly used in place of real people to pretest surveys, simulate users and communities, and evaluate AI systems. Each simulated person is defined by a persona, a written profile of fixed traits, demographics, and attitudes that a language model is instructed to act out. Every such system assumes that a person's identity is a stable attribute that behaves the same way in every situation the simulation presents. We show that this assumption is not a harmless simplification. Three decades of personality research find the opposite. How a given person behaves varies more across contexts than people vary from one another. Drawing on identity economics, which describes how different situations activate different identities and norms within the same person, we separate behavior into three parts, which are stable traits, the identity a context calls up, and the situation itself. We then prove, under two stated assumptions, that any simulation built on fixed personas, no matter how detailed the personas or how large the underlying model, cannot reproduce the part of a population's behavior that is specific to it and its circumstances. We propose a new evaluation criterion based on this result and a person-in-context design that can meet it in principle. We test one version on a naturalistic Reddit personality corpus and on a published clinical-trial recruitment survey of a community with documented distrust of medical institutions. The design does produce context-dependent behavior when tested directly, but in neither setting does it fully match the real population, and we show where it falls short. The consequence for practice is a risk that synthetic-population evaluations misrepresent the groups whose behavior is most shaped by context, a distortion that richer personas cannot correct.
Controlling the motion of multiple objects in image-to-video (I2V) generation requires preserving object identities while enforcing adherence to distinct target trajectories. This becomes particularly challenging as the number of objects increases and their paths intersect or occlude one another. Existing approaches entangle multiple trajectories within a shared, dense conditioning signal, making object-level correspondence difficult to preserve in crowded scenes. We depart from this paradigm and enforce a strict, per object spatial constraint that isolates instances independently. Our method, TrajLoc, achieves this directly within the attention layers by substituting the cross-attention weights of each object token with a Gaussian heatmap centered on its target location at every frame. The same per object token interface carries trajectory and depth through a learned embedding and preserves identity by encoding first frame appearance in place of an object token. Evaluations across six datasets, featuring up to 20 simultaneously controlled objects and out of distribution real world scenes, demonstrate that our method consistently improves both visual fidelity and trajectory adherence. Applied to two architecturally distinct backbones (CogVideoX 5B and WaN 2.1 14B), our approach achieves average gains of +4.3 dB PSNR and a 51% reduction in trajectory end point error compared to the strongest baselines. Project page: https://sela-omer.github.io/traj-loc/
An autonomous research laboratory could conduct an entire study, from formulating questions to preparing manuscripts. Artificial intelligence (AI) offers one route towards this goal. Recent AI systems take on several research roles, but may draw conclusions that the evidence does not support. Here, we present an autonomous research laboratory in which agents based on large language models assess one another's work and can stop or redirect investigations. They can generate and reject research questions, require separate theoretical checks and conduct experiments designed to challenge their claims. We demonstrate these capabilities by investigating ensembles of quantum spins in diamond. The agents derived the minimum number of calibration settings needed to predict responses to a defined family of microwave pulse sequences. They checked the result computationally and designed experiments to distinguish competing predictions. An initial experiment agreed with the theoretical predictions, but repeated reference measurements showed that the signal was drifting during data collection. The agents therefore treated the result as inconclusive and revised the measurement protocol without relaxing the acceptance criteria. The revised experiment met those criteria and supported the theoretical predictions. This study demonstrates that collaborating AI agents can select research questions, derive theoretical results and test their predictions experimentally.
The determinantal complexity of a polynomial is the order of the smallest matrix whose entries are polynomials of degree at most one and whose determinant equals the polynomial. A lower bound for the permanent in this measure is the algebraic form of the question whether the permanent is harder to compute than the determinant. Over the complex numbers, the best known lower bound grows cubically in the size of the permanent. It holds even when the permanent only has to be a limit of such determinants. I prove that the cubic lower bound holds over every algebraically closed field of odd characteristic $p$, with a constant that depends on $p$. Hence the cubic bound does not depend on characteristic zero. I also prove the bound for Pfaffians, which are square roots of determinants of skew symmetric matrices. Lower bounds for Pfaffians imply the same bounds for determinants. Further, I prove the bound, divided by $a$, for matrices whose entries are polynomials of degree at most $a$. All three results come from one lower bound for smooth hypersurfaces. I prove it in every characteristic by counting intersection points together with their multiplicities.
Understanding natural Earth surface processes requires integrating topography, imagery, hydrology, and human infrastructure while remaining robust to geographic variation, yet few benchmarks support both challenges. We introduce EarthScape, a multimodal benchmark for geologic mapping and surface-aware learning under geographic domain shift. EarthScape combines aerial imagery, LiDAR-derived elevation, multi-scale terrain derivatives, and contextual vector data into 37 co-registered input channels per sample, together with pixel-level geologic masks. Its geographically separated regions share a common label space, enabling controlled evaluation of in-domain and cross-region generalization under realistic class imbalance and spatial complexity. Benchmark experiments show that engineered terrain representations transfer more robustly than raw elevation, while spatial support and multimodal terrain inputs strongly influence performance. Fusion effectiveness depends on backbone and evaluation domain, with more complex cross-attention providing no consistent advantage over simpler strategies. These representation trends extend across geospatial foundation models and semantic segmentation, while class-level results show that no single terrain feature or scale is optimal across geologic classes. Together, EarthScape provides a benchmark for studying representation design, multimodal fusion, and geographic robustness in surface-aware geospatial learning. Our source code is available at https://github.com/masseygeo/earthscape and the dataset at https://doi.org/10.13023/kgs.data.05.01.2025.
Artificial intelligence offers statisticians a new paradigm for modelling and analysing complex datasets. Agents based on large language models can reason about data and automate key stages of the statistical modelling workflow, from model fitting to validation. In this paper we propose a method that uses such an agent to estimate the conditional density of a response given a vector of covariates from an iid sample. The agent examines the sample, chooses a model for the conditional distribution of the response and the procedure to fit it, and returns the estimator as a short computer program. The program is built from a library of probability distributions. The returned law can therefore be a parametric distribution, a mixture or a multivariate factorization, with distribution parameters given by explicit formulas in the covariates. We call this approach agentic distributional regression (ADR). Unlike in the classical approach, the statistician no longer chooses the estimator, but asks an agent to build one. Treating the returned program as a random conditional density estimator, we derive a finite-sample error bound under assumptions on the number of admissible models and on the accuracy of the agent's fit, and an oracle inequality for the selection of a model among fitted candidates using independent validation data. Finally, assuming that, as the sample size grows, the agent selects the same smooth parametric model with probability tending to one and fits it accurately, we show that the estimated parameters are asymptotically normal. In simulation studies, on twelve real datasets and on a new dataset whose responses postdate the agents' training data, ADR with Claude Opus 5.5 has the best average performance among the existing methods from the statistical literature with which it is compared.