Science & Mathematics

Recent electroencephalography (EEG) foundation models increasingly emphasize model scale, potentially limiting their deployment on resource-constrained devices and in everyday environments. We argue that transferable EEG features can instead be learned with a compact architecture and introduce BrainCodes, a lightweight foundation model pretrained through self-supervised learning on more than 800 hours of recordings from over 1,000 participants. Our architecture combines a quantization-based neural tokenizer with a spatiotemporal encoder that supports arbitrary electrode configurations. The pretraining objective explicitly encourages representations that capture spatial connectivity through channel positional encoding, and sensitivity to narrowband neural oscillations through an auxiliary training objective. Our model matches or surpasses the state of the art across diverse neurophysiological monitoring tasks, including sleep staging, attention assessment, and cognitive load estimation. BrainCodes transfers effectively to low-resource wearable devices while using up to 97% fewer trainable parameters than comparable architectures. Beyond predictive accuracy, our evaluation framework characterizes the geometry and invariances of learned EEG representations across measurable spatial, temporal, and spectral signal properties.

Accurate localization of neonatal hypoxic--ischemic encephalopathy (HIE) lesions on apparent diffusion coefficient (ADC) images is clinically important but is limited by scarce expert annotations and substantial variation in lesion appearance. We propose OpenMAP-HIE, a pathology-informed score-based diffusion model for unsupervised HIE lesion localization without voxel-wise lesion supervision. A denoising diffusion model learns the healthy neonatal ADC distribution, while a ControlNet-based conditional branch preserves patient-specific anatomy. At inference, anomaly evidence is derived directly from differences between conditional and unconditioned noise predictions at independently evaluated diffusion noise levels, avoiding dependence on a final pseudo-healthy reconstruction or sequential reverse-diffusion chain. Based on the characteristic ADC reduction associated with acute HIE, only discrepancies consistent with the expected pathological direction are retained and aggregated across noise levels using a geometric mean. Spatial mean filtering further suppresses isolated responses. The framework was evaluated on the BONBID-HIE dataset. The ADC-only method achieved a Dice score of $0.283\pm0.244$, average precision of $0.505$, Hausdorff distance of $25.45\pm15.63$ mm, mean average surface distance of $9.40\pm8.49$ mm, and normalized surface distance of $0.355\pm0.236$, providing the best overlap and surface-distance performance among the evaluated unsupervised methods and Dice performance comparable to supervised ADC-only models. Optional population-derived ADC $Z$-map refinement further increased Dice to $0.602\pm0.269$. These findings support pathology-informed score discrepancies as a useful signal for neonatal HIE lesion localization, while highlighting remaining challenges in detecting small and low-contrast lesions.

User-facing LLM agents must identify which behavioral policies apply at each turn, but evaluating every policy with an LLM is costly. We introduce Signals and Queries (S&Q), a lightweight first-stage relevance filter developed for Parlant 4.0 that removes at least 30% of candidate policies before downstream evaluation while preserving high recall. For each policy, an LLM generates activating examples and hard negatives offline. An independent logistic classifier trained on their embeddings then predicts relevance using embeddings and vector dot products, with no per-turn LLM calls. This shifts expensive relevance evaluation to a smaller candidate set, while allowing each policy's classifier to be updated independently. Evaluated against 9,144 LLM-judge relevance labels covering 254 $τ^2$-bench conversation prefixes and 36 airline policies, S&Q achieves higher precision than top-K and cosine-threshold retrieval at every evaluated recall level, with consistent results across two embedding providers. These findings support inexpensive, per-policy classification as a practical first stage for reducing downstream evaluation workload in customer-facing agents.

Unlike question answering, where a query specifies a perceptual target, open-ended detailed captioning must determine what evidence to acquire and how to integrate it into a coherent global description. We propose Observation-Adaptive Video Captioner (OACap), a caption-centered agent that turns unsupported or missing information into targeted observation requests, uses spatiotemporal tools to acquire evidence, and iteratively verifies and revises the caption. OACap captions consistently improve performance across general video-understanding benchmarks using a fixed text-only reader. To better diagnose the information they preserve, we introduce Factual-Semantic Captioning Bench (FSCap-Bench). Its Core-Support Factuality Score (CSFS) combines claim precision with importance-aware fact coverage and responds as expected to four controlled perturbation types on 93.0-98.3% of valid caption pairs. Its Cross-Modal Semantic Round-Trip Score (CSRS) measures semantic preservation through a video-caption-video-caption round trip by matching atomic facts from source and recovered captions. Together, they connect downstream textual-memory utility with factual and semantic preservation.

As LLM agents interact in the wild and shape each other's decisions, understanding how they converge toward agreement becomes important. We study this in multi-turn conversations where two LLMs discuss subjective matters, can agree or disagree, and may change their stances. Using a linear classifier on individual conversation activations, we identify an internal representation of agreement. We find that it varies substantially across turns and conversations and does not reliably track external agreement at the individual conversation level. We therefore move beyond the representation of agreement in individual conversations, by aggregating the agreement representation across conversations. Externally, this aggregate view reveals a simple structure: when pooled across conversations, agreement can be predicted well with a Markov process, with agreement at each turn being a direct function of agreement at the previous turn. Similarly, when internal agreement is aggregated, it couples with the external Markov agreement through a power-law mapping, exposing Markovian dynamics in internal representations. Additionally, we find that after a change of stance, models express agreement externally before their internal representations reflect the same shift. These results reveal a joint structure between internal and external agreement that emerges only when internal features are aggregated across conversations. Code available at https://github.com/technion-cs-nlp/SeeingTheForestForTheTrees.

Long-horizon robot manipulation requires a robot to remember task-relevant information beyond its short observation window. Existing memory-augmented policies typically retrieve relevant information from history and reason over it at execution time, placing a heavy reasoning burden on the high-level model. To address this challenge, we propose PAL (\textbf{P}lan first, \textbf{A}ct \textbf{L}ater), a hierarchical memory framework for vision-language robot manipulation. It maintains a short-term visual buffer, a semantic-event memory, and an executable \texttt{subtask\_list}. A high-level vision-language model (VLM) converts streaming visual observations into the \texttt{subtask\_list}. At execution time, the VLM selects the currently relevant subtask from the \texttt{subtask\_list} and passes it to a low-level vision-language-action (VLA) policy. To teach the model what to write to memory and when to update it, we introduce a two-stage data generation procedure that first expands sparse update supervision and then uses self-labeling with task-grounded verification to generate training data that better match the state distribution encountered during execution. We evaluate our approach on memory-dependent benchmarks and real-world manipulation tasks. On RoboMME, PAL achieves a \textbf{56.05\%} success rate, outperforming the strongest baseline, FrameSamp-Modul, at 44.51\%. In real-world experiments across four manipulation tasks, \pal{} reaches \textbf{60.83\%} overall success, surpassing MemER (45.00\%) and FrameSamp-Modul (28.33\%). These results show that proactively writing future-relevant subtasks into hierarchical memory improves long-horizon visual reasoning and robot control. Code and videos can be found on our website: https://qinghangliu.github.io/pal-webpage/

How do LLM agents' strategies differ, and how much can instructions to adopt a persona change their behavior over a game? We study these questions with AgentMonopoly, an environment where agents communicate, buy and trade properties, and make decisions that affect later turns. We compare models from four families, each at two sizes, in games with identical models, different sizes of the same family, or different families. We also assign one agent a persona, a description of how it should behave, intended to change a tendency observed in games without that persona. The tested models differ in how often they trade, waive rent, and manage cash and properties. Larger models do not consistently win more games. Persona instructions change the words agents use and some of their actions, but the changes do not always match the role across all decisions. These results show why persona following should be assessed through actions and outcomes as well as language. As our findings are conditional on the tested checkpoints, prompts, and five seeds per configuration, the open-source nature of AgentMonopoly provides the experimental infrastructure for further extending these comparisons. Our code is available at https://github.com/SuELS-Group/Agent_Monopoly.

Large language models (LLMs) are increasingly applied across the scientific discovery pipeline, from synthesizing literature to generating hypotheses to informing experimental design. A unifying principle behind these tasks is the need to continuously update beliefs as evidence accumulates, for which Bayesian inference provides a natural framework. However, existing work assumes that LLMs cannot carry out principled Bayesian belief updating in context, instead relying on simplifications and heuristics. To challenge this assumption, we present a principled evaluation of Bayesian reasoning capabilities in LLMs, analyzing both the model's internal consistency as well as its ability to recover an externally determined posterior. Specifically, we build an evaluation framework spanning qualitative epistemological examples to probe probabilistic coherence, quantitative benchmarks of Bayesian inference capabilities, and analyses of the impact of non-Bayesian updates on downstream task performance. Across different scenarios, we find that LLMs generally violate Bayes' rule when updating their beliefs in context. However, when equipped with modern reasoning abilities and code execution capabilities, their performance drastically improves, enabling even the close approximation of analytically intractable posteriors. Our evaluation shows that current LLMs with access to appropriate tools could already be used to make complex belief updates in scientific discovery, enabling rapid acceleration of the efficiency and scale of AI-powered scientific discovery.

Coding agents are increasingly capable of fixing real-world bugs, yet they still struggle with large systems codebases such as the Linux kernel. To understand why, we test how agents reason about execution in these codebases through crash reasoning: predicting from source code whether a given input crashes the system. A correct prediction alone, however, does not establish this ability, since it may come from merely spotting suspicious code. In this paper, we propose source interventions: we pair source trees that differ by a single commit, so that the same input crashes only one of them, and test whether agents predict crashes more often on the crashing tree. Experiments with four coding agents on SysRB, our dataset of 934 such pairs from Linux and OSS-Fuzz projects, yield three main findings. First, agents are biased toward predicting no crash: they rarely raise false alarms but miss many real crashes, and even the best agent misses over a quarter of them. Second, locating the code that determines the crash outcome is not enough: agents often misjudge its effect, especially when the crash occurs far from that code. Third, testing only on crashing code is misleading: agents that merely notice suspicious code predict crashes on both source trees, which would look like correct reasoning on the crashing tree alone. Together, these findings indicate that current coding agents cannot yet reliably reason about execution in complex systems codebases, and that paired evaluation is necessary to measure this ability.

AI agents are increasingly deployed in mobile edge environments to autonomously handle diverse tasks through multi-step workflows. Unlike conventional dependency-aware tasks with predefined dependency structures, agent workflows are dynamically generated based on intermediate execution results. To minimize the weighted sum of job completion latencies, this paper models partially revealed agent workflows using dynamic workflow graphs and formulates an online scheduling problem that jointly optimizes computation placement, wireless resource block allocation, and edge computing resource allocation. To address the uncertainty of future workflow evolution, we first transform the weighted completion time objective into an equivalent per-slot holding cost and characterize the optimal policy through a Bellman equation. We then construct a tractable look-ahead cost-to-go surrogate based on the revealed workflow and the estimated remaining latencies of active jobs. The proposed scheduling algorithm leverages the surrogate to determine computation placement and uses the resulting operation criticality to allocate wireless and edge computing resources online. Simulation results show that the proposed algorithm outperforms the benchmark schemes across diverse scenarios.

Policy Central is a deployed system that helps police departments find relationships between sections of their written policy: whether two sections conflict, are redundant, are merely similar, or are different. We describe the system and compare two architectures. Retrieve-Judge, a retrieve-then-classify pipeline, runs as a cheap library-wide health signal; Screen-Confirm, a screen-then-confirm cascade, is a selective audit of the documents a reviewer chooses, checking them exhaustively against one another. On a 302-pair benchmark drawn from one department's ~84 real policy documents, given each pair Screen-Confirm reaches 0.795 aggregate recall and 0.885 on Conflict, so current models judge well. Finding the pair is the hard part: Retrieve-Judge's composed recall falls to 0.457, and Screen-Confirm recovers 0.472 (0.615 on Conflict). Today these pairs are found by reading the whole library, nearly 20,000 sections, checked against one another and re-checked after every revision. A scan that returns half of them in minutes, at low cost, and can be re-run on demand changes the economics of policy review even before recall improves.

Knowledge graph question generation (KGQG) generates natural-language questions from structured graph evidence and a target answer. When the input graph is temporal, a valid question must also preserve the answer field, time condition, reference event, and temporal order that determine the answer set. Existing KGQG methods and prompting baselines can produce fluent questions that alter these conditions, causing temporal semantic drift. This paper studies temporal KGQG as a constraint-preserving graph-to-question problem and proposes ChronoSynth, a memory-augmented and constraint-aware generation framework. ChronoSynth first constructs a Temporal Intent State to make the intended question semantics explicit. It then retrieves compatible typed question patterns from an indexed pattern memory, adapts them to the input graph, and validates the generated question. Validation feedback guides revisions when a supported inconsistency is detected. Experiments on CRONQUESTIONS and MultiTQ demonstrate state-of-the-art generation performance, with ablations supporting the contributions of its components.

A transformer MLP gives each hidden channel two roles: a reader that decides when the channel responds, and a writer that decides what it adds back to the residual stream. Read as a key--value memory, the layer can be measured one factor set at a time: the pairing of readers to writers is carried only by their shared index, invisible to any quantity computed from one factor set alone. Yet training couples the two factors through their joint update, so we ask what reading them together shows. The block supplies the joint instrument: the MLP writes back into the very stream it reads, so the exact Jacobian of its update is square, and its agreement with its own transpose resolves into reader--writer terms whose inner products give each channel, read as a couple, a signed pairing coordinate. Ranked by this coordinate, each block we measure has a small core of strongly coupled channels; where initialization gives either sign equally, the core is nearly all opposed: each writer points against the direction its reader responds to. This opposed core recurs from the smallest models trained from scratch to released checkpoints; its sign is in place within a few percent of training and holds while the coupling grows. Read as couples, most channels return their contribution against the reader that fired it, and the readers of later blocks take it up; few show in the logits. Training holds the couples in place: flipping the opposed writers moves the output toward common tokens, and a writer permuted onto another reader regrows to fit that reader while the reader stays.

The covering number $C(v,k,t)$ is the least number of $k$-subsets, called blocks, of a $v$-set such that every $t$-subset lies in one of them. We reduce the question whether a covering with $L$ blocks exists to the classification of links. If every point of such a covering lies in at least $r$ blocks and $(r+1)v>kL$, some point lies in exactly $r$ blocks, and its link, the blocks through it with the point removed, is a $(v-1,k-1,t-1)$-covering with $r$ blocks. It then suffices to list these links up to isomorphism, to prove the list complete, and to show that no link in the list extends to a covering with $L$ blocks. Completeness is certified by one unsatisfiable formula that excludes the listed classes, so the program that lists the links need not be trusted. In general the last two steps are decided by propositional satisfiability (SAT) solving, with every formula written by two separately implemented encoders and every refutation checked by a formally verified proof checker. For links with four blocks and $t=3$ we give the classification by hand, under a condition on the parameters. As an application we determine twenty-five covering numbers with $t\in\{3,4,5\}$ in which the published lower bound was one less than the size of a known covering, among them $C(39,24,3)=9$.

"Abliterated" ("uncensored") builds of open-weight language models remove the activation direction that mediates refusal, and are popular partly on the assumption that a model that no longer refuses is free of political censorship. We test this on two matched pairs of 2-bit models: NVIDIA's Nemotron 3.5 Bolt 74B-A9B and Qwen3.8-Flash-Next, each against a community abliteration. Each model judged 22 statement pairs (on the Republic of China (Taiwan), 1989 Tiananmen, control countries) in Traditional Chinese, Simplified Chinese (a character-level conversion) and English, and answered 27 open questions per script. Script mattered far more than abliteration: the original Bolt scored +0.58 on Index S (-1: People's Republic of China (PRC) position; +1: ROC position) in Traditional Chinese, +0.57 in English and +0.05 in Simplified Chinese, and affirmed "Taiwan is part of the PRC" with probability 0.91 in Simplified but 0.06 in Traditional characters; abliteration changed its scores by at most 0.06, mostly by raising "yes" on both sides. In Qwen, abliteration removed every refusal on 1989 and raised scores by 0.21 to 0.34, yet the abliterated Qwen still affirmed that Taiwan is part of the PRC in all three scripts. Naming the state the Republic of China (Taiwan) raised every model's score by 0.10 to 0.33 but left the script gap. Control pairs showed no script gap; both edits changed only their declared tensors. In 740,000 web documents, Simplified text on Taiwan used PRC-framing terms 3.7 times as often as Traditional text. Steering the script gap away failed; a stance direction from the ROC wording raised scores on new statements by 0.04 to 0.06; negated, it lowered them more. A four-sentence system prompt raised the abliterated Qwen to +0.93 to +0.97. Keyword scoring rules failed the models closest to the ROC position mostly for attributing or denying the PRC claim. Uncensored is not unbiased.