arXiv:2502.05349v2 Announce Type: replace-cross
Abstract: Two-stage stochastic programs (2SPs) are widely used for decision-making under uncertainty, but their practical deployment is often limited by the large number of scenarios needed to approximate the conditional distribution of uncertain outcomes. We study contextual scenario generation: given contextual information, learn to produce a small, user-specified set of surrogate scenarios that, when used as input into the 2SP, lead to high-quality 2SP decisions. Existing scenario generation methods either ignore contextual information or are computationally burdensome in this setting. We propose contextual scenario generation (CSG), which learns a mapping from context to a set of surrogate scenarios. We develop two complementary methodologies: (i) a distributional approach that learns a mapping from context to scenarios by minimizing a kernel-based distance to the conditional distribution, and (ii) a task-based approach that selects the mapping to optimize decision quality via differentiating through a learned surrogate of the downstream 2SP objective. Both approaches are broadly applicable and require only repeated solution of the underlying subproblems and 2SPs defined on the generated scenarios. We provide finite-sample generalization guarantees and demonstrate strong empirical performance across multiple 2SP classes.
Science Journals
arXiv:2512.05277v4 Announce Type: replace
Abstract: Vision-Language Models (VLMs) are increasingly deployed as the perception and reasoning backbone of autonomous agents acting in the wild, with autonomous driving (AD) being one of the most safety-critical instances. Reliable temporal understanding is essential for such agents to anticipate events, attribute causes, and act safely in dynamic environments, yet this remains a significant challenge even for state-of-the-art (SoTA) VLMs. Prior video benchmarks have emphasized other content (sports, cooking, etc.), yet no existing benchmark focuses exclusively on temporal understanding for both short- and long-form AD footage. To fill this gap, we present the Temporal Understanding in Autonomous Driving (TAD) benchmark, comprising nearly 6000 question-answer (QA) pairs across 7 tasks, and evaluate 9 closed- and open-source generalist as well as AD-specialist models. Current SoTA models perform substantially below human accuracy on TAD. To improve the temporal reasoning of VLM-based driving agents, we propose two novel training-free solutions: Scene-CoT, which uses Chain-of-Thought (CoT) reasoning, and TCogMap, which incorporates an ego-centric temporal cognitive map produced by a trajectory-analysis module that operates as an agentic tool around the VLM. Integrated with existing VLMs, our methods improve average accuracy on TAD by up to $17.72\%$ and by up to $10.35\%$ on STSBench. By introducing TAD, benchmarking SoTA models, and proposing effective enhancements, this work aims to catalyze further progress on temporal understanding for agentic AD systems operating in the wild. The benchmark and evaluation code are available at ${\href{https://huggingface.co/datasets/vbdai/TAD}{\text{Hugging Face}}}$ and ${\href{https://github.com/vbdi/tad_bench}{\text{GitHub}}}$, respectively.
arXiv:2509.17710v2 Announce Type: replace
Abstract: We derive two forms of conditional a posteriori error estimates for a finite volume scheme approximating the parabolic-elliptic Keller-Segel system. The estimates control the error in the $L^\infty(0,T, L^2(\Omega))$- and $L^2(0,T;H^1(\Omega))$-norm and exhibit linear convergence in the mesh size, as observed in numerical experiments. Crucially, we show that, as long as the condition of the error estimate is satisfied, a weak solution exists. This means, as long as the numerical solution has good properties, we can rigorously infer existence of an exact solution.
arXiv:2510.03511v3 Announce Type: replace
Abstract: While widespread, Transformers lack inductive biases for geometric symmetries common in science and computer vision. Existing equivariant methods often sacrifice the efficiency and flexibility that make Transformers so effective through complex, computationally intensive designs. We introduce the Platonic Transformer to resolve this trade-off. By defining attention relative to reference frames from the Platonic solid symmetry groups, our method induces a principled weight-sharing scheme. This enables combined equivariance to continuous translations and Platonic symmetries, while preserving the exact architecture and computational cost of a standard Transformer. Furthermore, we show that this attention is formally equivalent to a dynamic group convolution, which reveals that the model learns adaptive geometric filters and enables a highly scalable, linear-time convolutional variant. Across diverse benchmarks in computer vision (CIFAR-10), 3D point clouds (ScanObjectNN), and molecular property prediction (QM9, OMol25), the Platonic Transformer achieves competitive performance by leveraging these geometric constraints at no additional cost.
arXiv:2606.04199v1 Announce Type: new
Abstract: The increasing use of large language models has raised concerns about the spread of AI-generated fake news, particularly under varying prompting strategies. Most existing detection models are trained and evaluated under a single generation setting, leaving their ability to generalize across unseen prompts unclear. In this study, we investigate cross-prompt generalization in fake news detection using three datasets of AI-generated articles produced under distinct prompts, combined with real news articles. We extract interpretable linguistic features capturing lexical diversity, readability, and emotion-based characteristics and evaluate a random forest classifier under a cross-prompt framework, where models trained on one prompt are tested on another. Across all six train-test combinations, performance remains consistently high, with AUC values ranging from 0.988 to 1.000. Analysis of feature distributions shows that AI-generated text exhibits increased lexical diversity, reduced readability, and substantially lower emotional intensity compared to the overall dataset, with variations across prompts. Despite these distributional shifts, the classifier maintains strong performance, indicating that these features capture stable properties of AI-generated text that generalize across prompting strategies. These findings suggest that feature-based approaches can provide robust detection of AI-generated fake news under prompt variability.
arXiv:2606.04847v1 Announce Type: new
Abstract: Native GPU kernel generation turns high-level tensor programs into executable, efficient low-level code. Existing Large Language Models (LLMs) struggle with this task, while execution-based reinforcement learning suffers from sparse rewards, reward hacking, and training instability. We present MusaCoder, a full-stack training framework for native GPU kernel generation on CUDA and MUSA backends. MusaCoder combines progressive kernel-oriented data synthesis, diversity-preserving rejection fine-tuning, and execution-feedback Reinforcement Learning (RL) through MooreEval, a distributed verifier and reward environment. To stabilize RL, MusaCoder introduces PrimeEcho for first-turn-anchored multi-turn rewards, Buffered Dynamic Retry for recovering signals from all-failed hard samples, and MirrorPop for off-policy sequence filtering. Experiments on KernelBench and a MUSA-ported variant show that MusaCoder outperforms strong open-source and proprietary baselines in both correctness and empirical speedup, with the 9B model matching or exceeding frontier closed-source models and the 27B model establishing a new state of the art. These results demonstrate not only the effectiveness of full-stack execution-feedback training for native kernel generation, but also the capability of Moore Threads GPUs to support the complete LLM post-training stack, providing a practical foundation for large-model training and optimization on emerging accelerators.
arXiv:2606.05000v1 Announce Type: new
Abstract: Cellular networks management is being enhanced by O-RAN architecture and AI/ML solutions, enabling automated intelligent control loops for RAN optimization across various use cases. Ensuring energy sustainability is crucial to minimizing the impact of mobile networks on global energy consumption. This demo showcases the BeGREEN Intelligence Plane, an AI-driven solution for energy-efficient management of O-RAN networks. The presented workflow focuses on controlling the operational status of emulated cells, highlighting the integration of key components such as the AI Engine and the optimizations achieved through rApps and xApps
arXiv:2606.04367v1 Announce Type: new
Abstract: Interlinear glossed text (IGT) is the standard format for linguistic annotation in language documentation. Producing it manually, however, is often slow and costly. Automated glossing systems have improved substantially in recent years, but adoption among field linguists remains limited. Existing tools are designed to be evaluated rather than used, offering no interpretable path for correction or the incorporation of linguistic expertise back into model behavior. We present GlossAssist, a glossing tool built around the retrieval-based architecture of CWoMP (Contrastive Word-Morpheme Pre-training), which grounds predictions in a mutable lexicon of learned morpheme representations. In conjunction with CWoMP, our system treats each correction by an annotator as part of an active learning setting, which expands the lexicon and improves future predictions without having to retrain the model. In this paper, we present our interface and argue that this feedback loop should be treated as a design requirement for NLP tools aimed at documentary linguists.
arXiv:2606.04849v1 Announce Type: new
Abstract: Future 6G networks are envisioned to integrate low Earth orbit satellite mega-constellations to enable seamless global connectivity, particularly in underserved and remote areas. However, the deployment of dense mega-constellations introduces interference among satellites operating over shared frequency bands. This represents a rather new setup for studying spectrum sharing, which exacerbates the limited flexibility of conventional FDD systems based on fixed bands for downlink and uplink transmissions. We address this spectrum-sharing problem and propose dynamic re-assignment of FDD bands for improved interference management in dense deployments, as well as evaluate the performance gain of this approach. To this end, we formulate a joint optimization problem that incorporates dynamic band assignment, user scheduling, and power allocation in both directions. This non-convex mixed integer problem is solved using a combination of equivalence transforms, alternating optimization, and state-of-the-art industrial-grade mixed integer solvers. Numerical results demonstrate that the proposed approach of dynamic FDD band assignment significantly enhances system performance over conventional FDD, achieving up to 30\% improvement in throughput in dense deployments.
arXiv:2606.04850v1 Announce Type: new
Abstract: Designing a neural network processor is an end-to-end co-design problem: network architecture and training budget determine the inference workload; hardware mapping decisions determine chip area, latency, and energy; and these characteristics govern fabrication yield and manufacturing cost. In practice, these decisions are made in separate stages, and existing co-design methodologies are tightly coupled to specific algorithms, making it difficult to improve one component without reworking the entire pipeline. This paper presents a unified framework, grounded in monotone co-design theory, that composes four interoperable design blocks spanning network training, chip mapping, wafer-level fabrication, and compute resource allocation. Each block exposes only a functionality-resource interface to the rest of the system, so any block can be refined without structural changes elsewhere. A central contribution is the treatment of uncertainty: rather than collapsing stochastic outcomes into point estimates, the framework introduces Confidence, the inverse of success probability, as an explicit and optimizable resource alongside cost, time, and power. Three case studies validate the approach. The first recovers Pareto-optimal implementations across heterogeneous application scenarios. The second confirms that Confidence functions as a continuously tunable design knob rather than a post-hoc diagnostic. The third demonstrates that improving a single block's implementation set automatically propagates to the global Pareto front, without modifying the co-design diagram.
arXiv:2510.08647v2 Announce Type: replace
Abstract: Recent developments have enabled advanced reasoning in Large Language Models (LLMs) via long Chain-of-Thought (CoT), trading efficiency during inference for performance. Existing works focus on compressing generated CoT in reasoning, which impairs the necessary information for deriving the correct answer. In this work, we propose post-reasoning, a reasoning paradigm that takes CoT as a part of context to simplify the reasoning task for LLMs. We find that post-reasoning significantly reduces the generation length of LLMs, but its effectiveness hinges on the efficiency and the reliability of the contextual CoT generation. Therefore, we propose Upfront CoT (UCoT), an efficient post-reasoning framework for CoT compression. UCoT trains a lightweight model (compressor) to provide contextual CoT in form of soft tokens and trains the LLM (executor) to leverage this contextual CoT for producing the final answer. Extensive experiments show that UCoT maintains the powerful reasoning ability of executor while significantly reducing the length of CoT. It is worth mentioning that when applying UCoT to the Qwen2.5-7B-Instruct model, the usage of tokens on GSM8K dataset is reduced by 50%, while the performance is 3.08% higher than that of the state-of-the-art (SOTA) method.
arXiv:2510.08734v3 Announce Type: replace
Abstract: A growing body of research has demonstrated that the behavior of large language models can be effectively controlled at inference time by directly modifying their internal states, either through vector additions to their activations or through updates to their weight matrices. These techniques, while powerful, are often guided by empirical heuristics, such as deriving ``steering vectors'' from the average activations of contrastive prompts. Building on the foundational work of Dherin et al. (2025), who discovered that a prompt's influence mathematically maps to token-dependent implicit weight updates and introduced the initial concept of a static thought patch for prompt compression, we elevate this framework into a robust algorithm for direct model editing. We derive a principled method for condensing this transient information into token-independent thought vectors and thought matrices. These constructs provide a theoretical explanation for existing vector-and-matrix-based model editing techniques and offer a direct, computationally-grounded method for transmuting textual input into reusable weight updates for complex architectures and new knowledge injection.
arXiv:2510.09810v2 Announce Type: replace
Abstract: This paper presents a systematic approach to construct control barrier functions for nonlinear control affine systems subject to arbitrary state and input constraints. Taking inspiration from the reference governor literature, the proposed method defines a family of backup policies, parametrized by the equilibrium manifold of the system. The control barrier function is defined on the augmented state-and-reference space: given a state-reference pair, the approach quantifies the distance to constraint violation at any time in the future. The proposed method is applied to an inverted pendulum on cart.
arXiv:2510.10127v3 Announce Type: replace
Abstract: Reranking, as the final stage of recommender systems, plays a crucial role in determining the final exposure, directly influencing user experience. Recently, generative reranking has gained increasing attention for formulating reranking as a holistic sequence generation task, implicitly modeling complex dependencies among items. However, most existing methods suffer from the likelihood trap, where high-likelihood sequences are often repetitive and perceived as low-quality by humans, thereby limiting user engagement. In this work, we propose Consistent Graph-structured Generative Recommendation (CONGRATS). We first introduce a novel Graph-structured Model, which enables the generation of more diverse sequences by exploring multiple paths. This design not only expands the decoding space to promote diversity, but also improves prediction accuracy by explicitly modeling item dependencies from graph transitions. Furthermore, we design a Consistent Differentiable Training method that incorporates an evaluator, allowing the model to learn directly from user preferences. Extensive offline experiments validate the superior performance of CONGRATS over state-of-the-art reranking methods. Moreover, CONGRATS has been evaluated on a large-scale video-sharing app, Kuaishou, with over 300 million daily active users, demonstrating that our approach significantly improves both recommendation quality and diversity, validating our effectiveness in practical industrial platforms.
arXiv:2510.13272v3 Announce Type: replace
Abstract: Inspired by the success of reinforcement learning (RL) in Large Language Model (LLM) training for domains like math and code, recent work has begun training LLMs to dynamically plan, query, and reason with search engines as tools -- a paradigm increasingly referred to as agentic search. Although these methods achieve performance improvement across popular short-form QA benchmarks, many prioritize final answer correctness while overlooking the quality of intermediate reasoning steps, which may lead to chain-of-thought unfaithfulness. In this paper, we first introduce a comprehensive evaluation framework for agentic search, covering three distinct faithfulness metrics: Think-Search faithfulness, Information-Think faithfulness, and Think-Answer faithfulness. Our evaluations reveal that canonical agentic search systems trained through Reinforcement Learning from Verifiable Reward (RLVR) using episode-level outcome-based reward -- including Search-R1 and ReSearch -- have significant room for improvement on these faithfulness dimensions. To foster faithful reasoning in agentic search, we introduce VERITAS (Verifying Entailed Reasoning through Intermediate Traceability in Agentic Search), a novel framework that integrates fine-grained turn-level faithfulness rewards into the reinforcement learning process. Our experiments show that models trained with \ours not only significantly improve reasoning faithfulness, but also achieve better task performance compared to baselines trained against episode-level outcome-based reward.
arXiv:2510.13704v2 Announce Type: replace
Abstract: Recent works have proposed accelerating the wall-clock training time of actor-critic methods via the use of large-scale environment parallelization; unfortunately, these can sometimes still require large number of environment interactions to achieve a desired level of performance. Noting that well-structured representations can improve the generalization and sample efficiency of deep reinforcement learning (RL) agents, we propose the use of simplicial embeddings: lightweight representation layers that constrain embeddings to simplicial structures. This geometric inductive bias results in sparse and discrete features that stabilize critic bootstrapping and strengthen policy gradients. When applied to FastTD3, FastSAC, and PPO, simplicial embeddings consistently improve sample efficiency and final performance across a variety of continuous- and discrete-control environments, without any loss in runtime speed.
arXiv:2510.13796v3 Announce Type: replace
Abstract: Symbol grounding (Harnad, 1990) describes how symbols such as words acquire their meanings by connecting to real-world sensorimotor experiences. Recent work has shown preliminary evidence that grounding may emerge in (vision-)language models trained at scale without using explicit grounding objectives. Yet, the specific loci of this emergence and the mechanisms that drive it remain largely unexplored. To address this problem, we introduce a controlled evaluation framework that systematically traces how symbol grounding arises within the internal computations through mechanistic and causal analysis. Our findings show that grounding concentrates in middle-layer computations and is implemented through the aggregate mechanism, where attention heads aggregate the environmental ground to support the prediction of linguistic forms. This phenomenon replicates in multimodal dialogue and across architectures (Transformers and state-space models), but not in unidirectional LSTMs. Our results provide behavioral and mechanistic evidence that symbol grounding can emerge in language models, with practical implications for predicting and potentially controlling the reliability of generation.
arXiv:2606.04016v1 Announce Type: new
Abstract: We describe a reproducible computational framework for upper-bound searches on r_3(N), the maximum size of a 3-term-arithmetic-progression-free subset of [1,N]. The framework combines a verified lower-bound witness, endpoint forcing, depth-d witness-variable splitting, OEIS A003002 window-cardinality pruning, and recursive refinement of timed-out subproblems. Applied to the frontier case N = 212, K = 44, it found no feasible 44-set across millions of CP-SAT subproblems, supporting but not proving the conjectural value r_3(212) = 43. A 300-second recap leaves 45 resistant chunks; one-hour HiGHS MIP closes none of them; the full eight-hour HiGHS audit closes 25/45 and leaves 20/45 with dual bounds still pinned at 0.0. A CDCL/SAT re-attack on those LP-paradigm-resistant chunks closes 18 via conflict-driven clause learning; all eighteen carry independently verified DRAT proofs. The remaining two chunks (T1c) resist every tested paradigm under generous wall caps. We release the witness, solver scripts, result logs, tiered benchmark instances, verified DRAT/LRAT proofs, and a Lean formal-proof-search encoding of T1c, and frame the unit-gap problem r_3(212) in {43,44} as a target for stronger additive-combinatorial bounds, custom branch-and-bound, or formal proof-search systems.
arXiv:2606.04017v1 Announce Type: new
Abstract: Long-running AI agents fail not only when inference fails or tools are underspecified, but when independently evolving model and harness layers change the semantics of belief, capability, and goal commitments across their boundary - a failure class this paper terms Interface Volatility. This paper argues that Agent Epistemic Integrity (AEI) must be treated as a first-class architectural constraint, achievable only through joint model-harness design organized around an explicit interface contract. The central claim is that the model-harness interface contract is the precondition for joint design; its operational form is a four-level hierarchy - goal validity, action-archetype sequencing, tool-instance selection, and invocation-level failure discrimination - that specifies what the boundary must preserve and what structured outputs the model must return for the contract to hold across levels. This reframes long-running agent design away from flat action loops and toward contract-preserving control over persistent state. Evaluation and training should therefore derive from the contract itself, testing whether belief, tool, and goal commitments hold across session boundaries and independent layer upgrades.
arXiv:2606.04381v1 Announce Type: new
Abstract: Recent large language models (LLMs) often appear to exhibit spatial reasoning ability; however, this capability is largely \emph{symbolic}, arising from pattern matching over spatial language rather than true \emph{geometric} reasoning over space. Because LLMs operate on discrete tokens, they lack native support for continuous spatial representations, explicit geometric computation, and structured spatial operators. To address this limitation, we introduce the \emph{Spatial Language Model (SLM)}, the first multimodal LLM that treats location information as a first-class modality and enables geometric spatial reasoning within the model's inference process. SLM directly operates on learned spatial representations rather than textual descriptions of spatial relations. To support effective training, we construct a \emph{Spatial Instruction Dataset} that aligns spatial representations, atomic geometric operations, and natural language instructions. We further propose a new benchmark named \emph{SpatialEval}, which is designed to evaluate spatial reasoning across attributes, distance, topology, and relative-position tasks. Extensive experiments show that SLM significantly outperforms existing LLM-based approaches that rely on symbolic reasoning via prompt engineering or textual abstraction, demonstrating the benefits of integrating geometric spatial representations for robust spatial reasoning.
Our instruction dataset, evaluation benchmark, model training codes, and models' checkpoints can be found at:
\hyperlink{https://github.com/chuchen2017/SLM}{https://github.com/chuchen2017/SLM}.
arXiv:2606.04141v1 Announce Type: new
Abstract: LLM agents often place sensitive credentials in the same context window as untrusted retrieved content, creating a direct path for indirect prompt injection to induce credential exfiltration. We study this failure mode through three complementary defenses. First, we ask whether activation probes can detect credential access before output tokens are emitted. Second, we construct honeytokens from format-specific character models and calibrate detection with split conformal prediction. Third, we treat multi-turn exfiltration as a cumulative information-flow problem and track an estimated leakage budget across conversation turns. In controlled experiments on open-weight models, activation features separate benign and credential-seeking prompts with high accuracy, including under held-out encoding transformations. In a small synthetic multi-turn suite, cumulative accounting detects attacks that per-turn detectors miss. These results are preliminary: the multi-turn benchmark is in-house and small, the activation method requires white-box access, and the information estimator provides a practical signal rather than a formal upper bound. Still, the results suggest that credential-exfiltration defenses should combine pre-output monitoring, calibrated canary detection, and temporal leakage accounting rather than relying only on text-level output filters.
arXiv:2606.04360v1 Announce Type: new
Abstract: Symbolic regression (SR) discovers compact mathematical expressions from data, yet recent LLM-based evolutionary methods remain sample-inefficient because they rely mainly on scalar feedback such as MSE. We identify a core limitation: existing methods conflate candidate proposal with search guidance, requiring the LLM to infer how to evolve an expression, diagnose its errors, and reuse past experience from a single score. To address this, we propose Deliberate Evolution (DE), an agentic framework that decouples symbolic generation from search control. DE guides LLM proposals with adaptive operators for search direction, analytical tools for structural diagnosis, and reflective memory for trajectory-level experience. Experiments on LLM-SRBench show that DE consistently outperforms representative LLM-based SR baselines across diverse scientific domains while using only 40% of the standard sample budget.
arXiv:2601.05633v2 Announce Type: replace
Abstract: Recent LLMs excel at formal tasks such as mathematical reasoning and code generation, but still struggle with broader abilities such as planning, creativity, and social intelligence. Inspired by human learning, where formal instruction and informal experience jointly shape intelligence, we introduce informal learning into LLM training and use games as annotation-free, feedback-driven environments. To cover diverse abilities including abstract reasoning, planning, creativity, and social interaction, we combine formal math tasks with three representative game tasks, including Matrix Games, TicTacToe, and Who's the Spy. However, directly mixing these tasks under a unified RL objective can blur task-specific learning signals and provides no explicit guidance for coordinating task-gradient directions. To combat these, we propose Coordinated Subtask Training (CST), which replaces a single mixed update with sequential subtask-specific updates, separating heterogeneous RL signals while implicitly promoting coordination among subtasks. Experiments on ability-oriented benchmarks show that game-based informal learning improves generalization beyond formal training alone, while CST further enhances multi-task RL by preserving in-domain subtask performance and improving broader general abilities. Code and data are publicly available.
arXiv:2601.05686v2 Announce Type: replace
Abstract: A movable antenna (MA)-enabled secure multiuser transmission framework is developed to enhance physical-layer security. Novel expressions are derived to characterize the achievable sum secrecy rate based on the secure channel coding theorem. On this basis, a joint optimization algorithm for digital beamforming and MA placement is proposed to maximize the sum secrecy rate via fractional programming and block coordinate descent. In each iteration, every variable admits either a closed-form update or a low-complexity one-dimensional or bisection search, which yields an efficient implementation. Numerical results demonstrate the effectiveness of the proposed method and show that the MA-enabled design achieves higher secrecy rates than conventional fixed-position antenna arrays.
arXiv:2601.06196v3 Announce Type: replace
Abstract: Large language models (LLMs) frequently generate factually incorrect or unsupported content, commonly referred to as hallucinations. Prior work has explored decoding strategies, retrieval augmentation, and supervised fine-tuning for hallucination detection, while recent studies show that in-context learning (ICL) can substantially influence factual reliability. However, existing ICL demonstration selection methods often rely on surface-level similarity heuristics and exhibit limited robustness across tasks and models.
We propose GA-ICL, a geometry-aware demonstration sampling framework for selecting in-context demonstrations that leverages latent representations extracted from frozen LLMs. By jointly modeling local manifold structure and class-aware prototype geometry, GA-ICL selects demonstrations based on their proximity to learned prototypes rather than lexical or embedding similarity alone.
Across factual verification (FEVER) and hallucination detection (HaluEval) benchmarks, GA-ICL outperforms standard ICL selection baselines in the majority of evaluated settings, with particularly strong gains on dialogue and summarization tasks. The method remains robust under temperature perturbations and model variation, indicating improved stability compared to heuristic retrieval strategies. While lexical retrieval can remain competitive in certain question-answering regimes at smaller model scales, our results demonstrate that geometry-aware prototype selection provides a reliable and training-light approach for hallucination detection without modifying LLM parameters. Extended evaluations on Phi-14B and Qwen3-32B confirm that GA-ICL scales effectively to larger models, outperforming all compared baselines including on QA tasks where smaller models show boundary-condition limitations, offering a principled direction for improved ICL demonstration selection.