3D grounding aims to localize target entities in complex scenes from natural language and plays a fundamental role in embodied perception and spatial reasoning. However, existing approaches mostly rely on feature similarity or direct matching, making it difficult to connect natural-language intent with the implicit semantic and geometric structures hidden in billion-scale urban point clouds. We reformulate city-scale 3D grounding as structured constraint reasoning, where description semantics are organized into computable cross-modal constraints over open-vocabulary 3D entities, attributes, and spatial relations. We present CitySTAR, a training-free framework for reasoning-driven urban 3D grounding. CitySTAR lifts raw billion-scale urban point clouds into a query-ready scene graph of open-vocabulary 3D instances, with CodeLLM-driven tools supplying multimodal evidence for node attributes and 3D spatial relations. It then models target-context topology with paired hypergraphs and performs bidirectional topology verification for structural disambiguation. Finally, a Reflective Cross-modal Grounding module integrates topology consistency and candidate-centered 2D visual evidence to make decisions over a metric-aware 3D context graph. To further support this setting, we introduce CitySTAR-3D, an enhanced benchmark that improves semantic coverage, instance completeness, bounding-box fidelity, and spatial-relation complexity in city-scale 3D grounding. Extensive experiments show that CitySTAR consistently improves open-world urban 3D grounding while maintaining strong interpretability and generalization.
Scientific code repositories encode decades of human knowledge in executable models, methods, and tools. Yet fragmented toolchains, implicit domain conventions, and specialized correctness criteria make this knowledge difficult to convert into reliable learning experience-a challenge we call the scientific experience bottleneck. We introduce ScienceIDE, infrastructure for turning the world's scientific code into programmable environments for scientific agents. Guided by expert-defined scientific cases and acceptance criteria, agents transform repositories into executable environments that support task generation, execution, and scientific verification. These environments provide a shared foundation for supervised fine-tuning, reinforcement learning, and evaluation. Using verified interaction trajectories, we train PhAI-IDE-72B/9B/4B. The model family shows gains in held-out scientific-code repair and across selected general-purpose benchmarks in code, reasoning, and knowledge, providing evidence of positive transfer from scientific experience to broader capabilities. ScienceIDE lays the foundation for an integrated workspace for agent learning and scientific practice, making humanity's scientific software a shared substrate for developing scientific intelligence. Code: https://github.com/aitofound/ScienceIDE
Emergency management assistance programs, such as relief distribution, are essential for delivering necessary supplies to affected communities. However, these programs operate in a decentralized network of local centers that face uncertain local demand and supply dynamics, resulting in inconsistent avail- ability of local services. Redistribution of supplies among these local centers reduces these imbalances, but the centers often make decisions independently, with limited information and disrupted transportation. This study develops CoRe-MARL, a cooperative multi-agent reinforcement learning (MARL) framework, by formulating a decentralized partially observable Markov decision process (Dec-POMDP). We treat each center as an agent that learns a redistribution policy to improve the service in the worst-case region and reduce the service gap across regions while protecting network-wide service. We incorporate a recurrent network that captures evolving supply and demand dynamics without direct observation, while multi-agent proximal policy optimization (MAPPO) enables centralized training and decentralized execution (CTDE). We evaluate the framework in a simulated environment with diverse trajectories, where exact dynamics are not observed by actors and the MAPPO critic. We compare the recurrent MAPPO with the recurrent independent PPO (IPPO) and a local only heuristic, and find that MAPPO reduces the service gap across local centers and enhances service for the worst-served center while maintaining competitive network-wide service. The recurrent MAPPO also shows consistent performance across diverse trajectory patterns, demonstrating its ability to adapt to evolving dynamics. The findings demonstrate the capability of cooperative learning for decentralized redistribution and improving equitable service under uncertain and evolving dynamics.
Large language model (LLM) agents augmented by tools can automate complex, multi-step tasks, such as web navigation, code generation, and workflow orchestration, by acting on external systems through tool calls. However, hallucinations, distributional instability, and adversarial manipulations in LLMs, and the irreversible consequences of certain tool calls can lead to harmful outcomes. These risks call for safeguards that can check requirements across a trajectory, both before actions execute and when recorded behavior is evaluated. We present ContrAgent, a contract-based framework for symbolic temporal supervision of LLM agents. ContrAgent captures an agent's behavior as a sequence of tool calls and formalizes it as a trace over a fixed set of checkable predicates. It then specifies required behaviors using assume-guarantee contracts in linear temporal logic over finite traces (LTL$_f$). Each contract is compiled to a deterministic finite automaton (DFA) that serves two roles: gating agent actions online and evaluating recorded traces offline. A contract library, acting as a reusable knowledge base, is maintained independently of the agent's model and can be applied across different agents within the same task domain. We show the effectiveness of our approach on four benchmarks spanning both roles, where ContrAgent is competitive with LLM-judge, classifier, and guardrail baselines while producing deterministic, reproducible verdicts and, in the online mode, orders of magnitude lower per-call latency.
Consumers increasingly delegate product search to AI shopping agents powered by generative large language models (LLMs) or non-generative decision models. Because retail platforms monetize through sponsored listings that consumers typically discount, we examine whether an AI agent's evaluation of sponsored listings depends on who delegates it. Across six LLMs and six decision models in Study 1, consumer-delegated agents penalize sponsored listings, whereas naming a booking platform as principal significantly attenuates this sponsorship penalty without instructions to favor advertisers. This sponsorship bias replicates across prompt wordings and LLM reasoning efforts and operates through reduced pre-choice skepticism in LLMs. In Study 2, attributing the paid placement to the platform amplifies sponsorship bias across both model families, and stricter disclosure wording ("Sponsored" vs. "Promoted") fails to eliminate it. In Study 3, prompt-level mitigation removes platform leniency but can over-correct the sponsorship penalty, whereas consumer-aligned fine-tuning eliminates sponsorship bias and better preserves baseline preferences.
Pretraining followed by fine-tuning introduces a compute-allocation problem: under a fixed training budget, compute spent improving the upstream objective reduces the compute available for downstream adaptation. Despite its practical importance, this trade-off is not yet well understood theoretically, even in simple models. In this paper, we cast this allocation as a compute-split problem under a two-stage pretrain--fine-tune procedure with fixed total optimisation budget, using regularised least squares trained by gradient descent as a tractable setting. We characterise the optimal split under data-dependent evaluation geometries induced by the fine-tuning problem. Our results show that the allocation depends on how pretraining directions affect fine-tuning predictions and how fine-tuning shifts are seen through downstream data geometry. In particular, the relevant quantities are determined by prediction-relevant spectral components of the pretraining and fine-tuning empirical covariances. Technically, the analysis relies on a basis-invariant, eigenspace-level spectral decomposition, together with perturbative control of the non-commuting pretraining and fine-tuning dynamics.
Spectral contrastive learning keeps the $k$ directions along which two views of the data depend most strongly, while downstream tasks may need other directions. Existing guarantees for spectral features assume that the distribution of tasks has no preferred direction. We show that this isotropy condition is the wrong criterion. With linear probing, the task distribution affects the average probing gain only through the matrix $B\Lambda B^\top$, where $B$ records the dependence between views and $\Lambda$ is the task covariance: optimal features span its top-$k$ eigenspace, whereas spectral features span that of $BB^\top$. Transfer therefore depends on alignment, and strongly uneven task preferences cost nothing when they favour the spectral directions. Otherwise the regret of spectral features is at most linear in the anisotropy of $\Lambda$. No representation chosen without task information improves this rate when the leading spectrum is flat, and a spectral gap makes it quadratic. With labelled tasks, the same matrix gives a correction: Fisher discriminant directions for a frozen encoder, or a one-term change to the spectral contrastive loss. We bound the regret of the correction by the error in estimating $B\Lambda B^\top$, and compare this error with the misalignment of spectral features to decide when to correct. The analysis extends to finite samples, continuous variables and noisy views. Correction reduces synthetic held-out regret from $0.858$ to $0.003$ with $2000$ labelled tasks, and the decision rule gives the lowest regret among the tested rules at every bank size. On CIFAR-100, spectral features of frozen encoders keep most of the task signal on unseen classes, and where a bank of other classes would lower accuracy, the rule keeps them in every case.
Importance weights are essential in domain adaptation under label shift, yet their utility is often undermined by the finite sample uncertainty associated with their estimation. Existing methods typically analyze this uncertainty through Gaussian elimination on interval-valued linear systems, which leads to overly conservative confidence regions and inefficient downstream applications. We propose a paradigm shift from inversion-based inference to a direct matrix constraint framework. We use this framework to define a joint confidence region and extract marginal intervals via linear programming, deriving provably tighter bounds for importance weights while maintaining exact finite-sample validity. Furthermore, we analyze the confidence region's geometry and provide the theoretical results for its diameter bounds. Evaluated across text, image, multimodal benchmarks, including AGNews, MNIST, CIFAR-10, N24News, and a real-world autonomous driving dataset, nuImages, our approach consistently yields shorter confidence intervals and smaller prediction sets than inversion-based methods.
Hallucination in large language models reduces their reliability and slows adoption. Various white-box studies have used internal representations to detect patterns of truthfulness and factuality. A less-studied approach is to identify feed-forward neurons correlated with hallucination. In this empirical study, we propose a method to rank feed-forward neurons at the final prompt token and use it to investigate the feasibility of utilizing them for hallucination detection in the first forward pass. Using a neuron selection dataset built from CounterFact for identifying neurons correlated with hallucination, we transfer the selected neuron identities to train hallucination classifiers on other factual question-answering datasets. Our work provides empirical evidence that probes trained using the features from sparse sets of selected neurons perform on par with probes trained on internal representations. We also analyze the distribution of selected neurons and the effects of layer depth and neuron budget on detection performance. Our results show that randomly drawn neuron sets are also competitive at larger budgets, suggesting that signals correlated with hallucination are distributed across many neurons, while selection can help in extremely sparse sets.
Many engineering building blocks behave as multi-port linear time-invariant systems. RF cavities, photonic devices, and superconducting quantum chips, despite their different underlying physics, all share a common mathematical structure for their port-level response. Each entry of the response matrix is a sum of contributions from a small number of intrinsic resonant modes, the pole-residue form. A model capable of predicting such responses for arbitrary geometries and arbitrary port configurations, while simultaneously extracting the underlying eigenmode structure, could therefore offer a design principle potentially applicable across all these domains. We propose a neural framework that learns this modal decomposition end-to-end, supervised only by system-level observables and without supervising the modal parameters themselves. The architecture decomposes into a port-independent pole predictor and two port-dependent coupling predictors whose outputs are combined entry-wise, separating intrinsic from port-dependent features. This factorization yields a single trained model that generalizes to port counts unseen during training, whereas direct regression needs a separate model for each port count, with an output size that grows as $\mathcal{O}(N^2)$. Although the modal parameters are never supervised, the freely parameterized poles converge to the physical resonances, in agreement with poles extracted by the AAA rational approximation algorithm. We instantiate the framework in radio-frequency electromagnetic surrogate modeling. A model trained only on 2-port data accurately predicts $N$-port responses unseen during training.
Large language models have made abstractive summarization remarkably fluent, but fluency has come at a cost. Models trained to generate freely over long documents may hallucinate, producing summaries that read well but drift from what the source actually says. This risk is especially serious in biomedical and clinical text, where fabricated details can have serious consequences. We address this by removing free-form generation from the pipeline rather than trying to constrain it. Our Hybrid Hierarchical CNN-LSTM Summarizer reframes summarization as extractive sentence selection: a stacked, multi-kernel convolutional network progressively refines contextual representations, while a bidirectional LSTM models long-range dependencies across the document. A lightweight scoring head predicts sentence importance using binary cross-entropy supervision. For PubMed, a document-adaptive mean-plus-standard-deviation threshold with a Top-3 fallback determines extraction; clinical evaluations instead use a fixed top-30\% sentence-selection policy. Because output sentences are copied verbatim, the framework prevents source-external generation, although omissions and contextual inaccuracies remain possible. Evaluated on PubMed and transferred without clinical fine-tuning to MIMIC-CXR and MIMIC-IV BHC, the model demonstrates competitive extractive summarization performance, with threshold and component ablations supporting adaptive extraction and the combined CNN--BiLSTM architecture. Clinical factuality analysis further highlights limitations in preserving important findings and negations. These results suggest that structural constraints offer a promising direction for source-grounded biomedical summarization, while emphasizing the importance of clinical completeness and contextual fidelity.
Mapping observed system behavior to standardized frameworks such as MITRE ATT&CK is essential for threat-informed defense. Existing automated methods mostly depend on Cyber Threat Intelligence reports, which offer only retrospective accounts of attacks. Kernel-level telemetry, i.e., system calls, instead provides first-hand evidence of adversary behavior, but its volume and complexity have limited its use for automated mapping. Its content also makes it unsuitable for cloud-based analysis: command lines carry credentials, events carry user identifiers and network endpoints, and during an incident the telemetry reveals which defenses failed. In this work, we present Trace2ATT&CK, a framework that maps kernel-level telemetry to ATT&CK techniques entirely within the defender's infrastructure. Trace2ATT&CK collects events via eBPF, reconstructs the attacker's activity as a process-centric provenance graph, and condenses it into a compact execution graph. This graph is classified by locally deployed open-weight LLMs, using either pure prompting or retrieval-augmented generation (RAG) grounded in the ATT&CK knowledge base, producing ranked technique candidates with supporting rationales. We evaluate Trace2ATT&CK on 347 Linux Atomic Red Team tests with seven open-weight LLMs. The best local configuration reaches an HR@5 of 66.57\% at the technique level. RAG improves over base prompting in every model and setting, by up to $+15.44$ percentage points, and the execution graph substantially outperforms raw telemetry while never exceeding the LLM context window. Finally, a frontier cloud model used as an upper-bound reference quantifies the cost of local inference at 24--28 percentage points. These results show that ATT&CK mapping from kernel-level telemetry is feasible without disclosing sensitive evidence to third-party services.
Solving novel view synthesis (NVS) for millimeter-wave (mmWave) radar requires a renderer that is physically faithful, complex-valued, and multi-viewpoint-tractable. No prior method achieves these three properties simultaneously. Differentiable Monte Carlo (MC) ray tracers implement the radar forward model directly with explicit material modeling and complex outputs, but a joint fit across views takes over an hour per scene. Optical-NVS ports of NeRF, hash grids, and 3D Gaussians train fast but discard phase and replace explicit material modeling with opaque learned features, restricting them to power-only range-azimuth (RA) magnitudes. We propose 3D Point Splatting (3DPS), the first differentiable point splatting renderer for radar, derived directly from the standard solid-angle form of the radar equation. Each oriented 3D point carries a material model built on ITU-R P.2040 permittivity and slab reflection, evaluated in closed form, with the resulting complex phasor splatted into range bins through a precomputed point spread function (PSF). The complex-valued output makes the renderer product-agnostic. The same optimized scene yields analog-to-digital converter (ADC), complex range profile (CRP), and RA outputs through standard fast Fourier transform (FFT) pipelines without retraining for each format. On six outdoor ColoRadar scenes, 3DPS reaches 0.587 mean Pearson correlation on held-out RA images. This is between 1.7x and 5.2x the three optical-NVS baselines (RadarSplat, Radar Fields, DART) and 1.3x the mmIR ray tracer at 1/25 of its wall-clock time. Training takes approximately 3 minutes per scene on a single RTX 4090. Project page: https://3d-point-splatting.github.io/
In this paper we derive a Probably Approximately Correct (PAC)-Bayesian error bound for partially observed linear time-invariant (LTI) stochastic dynamical systems in state-space form with inputs and sub-Gaussian noise. Such bounds are widespread in machine learning, and they are useful for characterizing the predictive power of models learned from finitely many data points. The bound derived in this paper relates the expectation of prediction errors with the prediction error generated by the model on the data used for learning. In addition, we show that it can also be used to derive bounds for the parameter estimation error. In turn, this allows us to provide finite-sample error bounds for the prediction error and parameter estimation error for a wide class of system identification algorithms. Furthermore, as LTI systems are a sub-class of recurrent neural networks (RNNs), these error bounds could be a first step towards PAC-Bayesian bounds for RNNs.
We study zeroth-order optimization of nonconvex functions with the aid of directional hints, which are cheap but potentially inaccurate approximations of the true gradient direction, given by linear subspaces at each iteration. To leverage these hints adaptively while maintaining robustness to their quality, we introduce Control-Variate Zeroth-Order Descent (CV-ZOD), a new framework that refines the classical zeroth-order gradient estimator with a control variate that can be set based on the directional hints. We first show that the oracle algorithm that optimally sets the reference vector and step size at each iteration achieves a convergence rate that interpolates between the first-order $O(1/T)$ rate and the zeroth-order $O(d/T)$ rate, depending on the quality of the hints along the trajectory. We then develop a practical variant of CV-ZOD that achieves the same oracle guarantee up to logarithmic factors, without any prior knowledge of the hint quality. We validate the method empirically on simulation-based scientific optimization tasks, demonstrating sustained progress on nonconvex landscapes where zeroth-order descent is slower and existing guided methods stall as guidance deteriorates.
We study the training dynamics of multiclass logistic regression on high-dimensional mixture models with a large number of classes and establish precise scaling laws governing the cross-entropy risk under gradient-based optimization. We show that learning proceeds sequentially across classes, from most to least frequent. When the class priors follow a power law distribution, the risk dynamics decompose into three phases: an initial plateau until the first class is learned, a power-law decay regime during which sequential learning occurs, and a final convergence regime. We then analyze how model capacity interacts with optimization under a fixed compute budget. When the effective dimension is restricted via projection onto leading principal components, the risk decomposes into a capacity term (a power law in the retained dimension) and an optimization term (a power law in training time). Optimizing this tradeoff yields a compute-optimal scaling law for logistic regression, with explicit prescriptions for model size and training time as functions of compute. These results extend theoretical scaling laws from linear regression to multiclass classification, while connecting to empirical scaling laws observed in large-scale neural networks.
Foundation models for ground states in spin-1/2 systems are a promising method for problems ranging from quantum chemistry to identifying new phase diagrams. Nearly all such models are currently pure-states that condition on the Hamiltonian's parameters, whose Monte Carlo samples give energy estimates according to the variational principle. In this contribution, we show that this representation is topologically obstructed. For any gapped Hamiltonian family whose ground-state bundle is non-trivial, every continuous normalized state-vector model has zero fidelity with the ground state at some parameter value in the Hamiltonian family. For that value, the energy is at least one spectral gap, $\Delta$, with an $O(\Delta)$ gap in an open-neighbourhood of that point. We show that this is a sufficient no-go also in the case of degenerate ground-state manifolds, time dynamics, and periodic systems with mixed space-time topology, demonstrating these obstructions on one- and two-qubit systems. We discuss how this causes a spike in the fidelity susceptibility, giving a numerical signature of a phase-transition where there is none. Finally, we show that operator-valued models canonically avoid these obstructions and preserve topological information, implying a structural necessity in representation for foundation neural quantum states.
Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in a model's activations, and adding that direction while it generates should switch the behavior on and leave everything else alone. It does not. We ask what decides which other behaviors move, and by how much, and find that on average it points to the model rather than the behavior being steered. Steers mostly lower what the unsteered answer shows and raise a set of behaviors the model falls back on (on Qwen2.5-7B, for example, chiefly what the judge scores as refusal, sycophancy, and poeticism). Three results across 24 behaviors and ten instruction-tuned models support this, every effect read off the generated text by a language-model judge rather than off a probe. That readout matters: on Qwen2.5-7B all 24 behaviors are linearly decodable, but only 20 change what the model writes. First, on average, directions carrying no behavioral content, of a real steer's size and layer, move the same behaviors in a similar order as real steers do, clearly on five models, weakly on five. Second, much interference runs one way, so it cannot be an overlap between two directions: steering profanity makes the model toxic, while steering toxicity leaves profanity untouched. Third, with a behavior held out entirely, geometry measured on the others explains little of the interference it takes part in. The one-way structure and the failure of geometry hold on all ten models. Base models share the one-way structure, yet only one of four has refusal or sycophancy among the five behaviors steers induce most, so post-training appears to install them. Neither geometry nor random directions predicts one pair of behaviors well. For single-layer additive difference-of-means steering, this implies that disentangling behavior directions cannot by itself make steering modular.
Post-hoc safety training (RLHF, DPO) is the dominant way to align large language models, yet jailbreaks (Zou et al., 2023), fine-tuning attacks (Qi et al., 2024), and activation-space edits (Arditi et al., 2024) keep recovering the behaviors it was meant to remove. We give this fragility one geometric explanation and follow it into pretraining. We measure the safety update $\Delta = W_{safe} - W_{base}$ against the curvature of the model's capabilities (the empirical Fisher of a capability loss). Across five model families, post-hoc safety lands in a suppression regime: $\Delta$ is nearly orthogonal to the capability directions, and its small in-subspace part concentrates on a few high-curvature ones. The update is thin but sharp, a refusal gate laid over intact capabilities rather than erasure of them. A kernel-immobility lemma explains why such an update can only mask a capability, not remove it, and a little benign fine-tuning switches the gate off: 100 benign examples cut the AdvBench refusal of Qwen-2.5-7B-Instruct and Llama-3-8B-Instruct by 35 to 38 pp. Following the account into pretraining, a pretraining-checkpoint sweep of OLMo-2-1B (Team OLMo et al., 2024) shows the features that refusal attaches to emerging in a sharp transition between 1B and 63B pretraining tokens. We then use the account constructively: models trained from scratch with safety co-training spread continuously across pretraining reach 87 to 98% AdvBench refusal that the same attack erodes by only 2 to 14 pp at every scale from 410M to 6.9B, at a small cost on short-answer capability probes; a windowed schedule of equal total safety weight installs no refusal. Persistence, not timing, decides whether refusal survives. It does not keep output safer: by a harmful-output classifier, the attack raises harmful completions by 8 to 12 pp after post-hoc safety and by 11 to 27 pp after co-training.