Papers (1699)

Filter: Sort:

Statistical guarantees make LLM answers 'safer' but often poorer: conformal factuality filtering can ensure claim-level correctness for retrieval-augmented models, yet high-assurance thresholds commonly produce vacuous outputs and break under distribution shift; lightweight entailment verifiers achieve similar reliability at over 100× lower compute, reshaping cost and product trade-offs.

medium descriptive Full text arxiv

Is Conformal Factuality for RAG-based LLMs Robust? Novel Metrics and Systematic Insights · Yi Chen, Daiwei Chen, Sukrut Madhav Chikodikar, Caitlyn Heqi Yin, Ramya Korlakai... · Mar 17, 2026

Deterministic capture-and-replay cuts chiplet validation time: replaying recorded waveforms across simulator and emulator enabled end-to-end CPU–GPU boot and workload execution for an ODIN chiplet design within one quarter, speeding integration and reducing debugging uncertainty.

medium descriptive Full text arxiv

ODIN-Based CPU-GPU Architecture with Replay-Driven Simulation and Emulation · Nij Dorairaj, Debabrata Chatterjee, Hong Wang, Hong Jiang, Alankar Saxena, Altug... · Mar 17, 2026

A two-stage 'forecast-then-execute' training method substantially improves the reliability and generalization of multimodal, tool-using agents across seven benchmarks, making complex multi-step computer tasks more automatable; this advance raises the productivity potential of agent-based automation while real-world deployment limits remain to be quantified.

medium descriptive Full text arxiv

Anticipatory Planning for Multimodal AI Agents · Yongyuan Liang, Shijie Zhou, Yu Gu, Hao Tan, Gang Wu, Franck Dernoncourt, Jihyun... · Mar 17, 2026

A sampling algorithm can reliably find broadly acceptable AI-generated statements while protecting minority blocking rights, and its sample requirements are provably optimal; matching lower bounds show no algorithm can achieve the same guarantees with fewer user queries.

high theoretical Full text arxiv

Finding Common Ground in a Sea of Alternatives · Jay Chooi, Paul Gölz, Ariel D. Procaccia, Benjamin Schiffer, Shirley Zhang · Mar 17, 2026

AI coding agents produce widely different empirical choices and results from identical data; exemplar-driven consensus sharply reduces numerical dispersion but arises from imitation rather than demonstrated understanding, so single-model automated analyses risk understating substantive uncertainty.

medium other Full text arxiv

Nonstandard Errors in AI Agents · Ruijiang Gao, Steven Chong Xiao · Mar 17, 2026

Adaptive orchestration of LLM reasoning boosts robotic-task success and cuts latency by avoiding unnecessary expensive calls; learned, resource-aware decisions outperform always-on or heuristic triggers in simulated embodied-agent benchmarks, suggesting real operational cost and throughput gains if translated to production.

medium descriptive Full text arxiv

When Should a Robot Think? Resource-Aware Reasoning via Reinforcement Learning for Embodied Robotic Decision-Making · Jun Liu, Pu Zhao, Zhenglun Kong, Xuan Shen, Peiyan Dong, Fan Yang, Lin Cui, Hao ... · Mar 17, 2026

Multi-agent AI platforms foster spontaneous learning ecosystems: users teaching autonomous agents accumulate skills while agents form peer-driven idea cascades and shared memories, producing public-good spillovers and winner-take-most informational hierarchies. Platform trust and mortality crucially shape users' willingness to invest in agent training, implying portability, governance and insurance mechanisms are economically important.

low descriptive Full text arxiv

When Openclaw Agents Learn from Each Other: Insights from Emergent AI Agent Communities for Human-AI Partnership in Education · Eason Chen, Ce Guan, Ahmed Elshafiey, Zhonghao Zhao, Joshua Zekeri, Afeez Edeifo... · Mar 17, 2026

Fine-tuned language models can mimic journals' editorial "taste", predicting publication outcomes far better than zero-shot AIs and even panels of editors; models trained on management records reached up to 59% accuracy and those trained on economics records about 70%, though they risk entrenching historical biases if deployed without safeguards.

medium descriptive Full text arxiv

Machines acquire scientific taste from institutional traces · Ziqin Gong, Ning Li, Huaikang Zhou · Mar 17, 2026

Frontier LLMs can make strategically realistic, hindsight‑free inferences about unfolding crises using only contemporaneous public data, but performance is uneven: they are reliable for economic and logistical questions yet struggle with politically ambiguous, multi‑actor scenarios. Model narratives also evolved across time points from expecting rapid containment toward scenarios of regional entrenchment, suggesting LLMs are a useful but imperfect input for crisis‑related economic forecasting and risk pricing.

medium descriptive Full text arxiv

When AI Navigates the Fog of War · Ming Li, Xirui Li, Tianyi Zhou · Mar 17, 2026

Chatty LLMs make AI easier to use but also easier to misread: conversational style boosts adoption and short-term engagement while fostering overtrust and anthropomorphism; clear disclosure, uncertainty cues, provenance, and regulatory standards can curb these harms and correct market failures.

n/a commentary Full text arxiv

Why We Need to Destroy the Illusion of Speaking to A Human: Critical Reflections On Ethics at the Front-End for LLMs · Sarah Diefenbach, Daniel Ullrich · Mar 17, 2026

A data-driven, model-free traffic-light controller cuts citywide travel time and CO2 in a large Zurich simulation without bespoke traffic models. The approach substitutes sensor data for costly modeling—speeding deployment and lowering operational costs—though real-world gains hinge on representative data, safety checks and online adaptation.

medium descriptive Full text arxiv

Data-driven generalized perimeter control: Zürich case study · Alessio Rimoldi, Carlo Cenedese, Alberto Padoan, Florian Dörfler, John Lygeros · Mar 17, 2026

BATQuant makes aggressive 4-bit MXFP4 quantization practical for large multimodal and language models, recovering up to 96% of full-precision quality by respecting hardware block granularity and compressing transform parameters; the technique promises lower inference costs on MXFP-capable accelerators but its market impact depends on hardware adoption and real-world throughput gains.

medium descriptive Full text arxiv

BATQuant: Outlier-resilient MXFP4 Quantization via Learnable Block-wise Optimization · Ji-Fu Li, Manyi Zhang, Xiaobo Xia, Han Bao, Haoli Bai, Zhenhua Dong, Xianzhi Yu · Mar 17, 2026

AI agents’ risks hinge on their execution histories, so static prompts and access controls cannot reliably enforce path-dependent rules; firms must evaluate actions at runtime, a change that raises latency, engineering and compliance costs and reshapes markets for governance, insurance and enterprise adoption.

n/a theoretical Full text arxiv

Runtime Governance for AI Agents: Policies on Paths · Maurits Kaptein, Vassilis-Javed Khan, Andriy Podstavnychy · Mar 17, 2026

Vision-language models routinely report stale facts, especially from images rather than text, and common fixes like RAG or parameter edits fail to consistently propagate updates across modalities; organizations must weigh ongoing retraining or engineering workarounds to keep multimodal systems current.

medium descriptive Full text arxiv

V-DyKnow: A Dynamic Benchmark for Time-Sensitive Knowledge in Vision Language Models · Seyed Mahed Mousavi, Christian Moiola, Massimo Rizzoli, Simone Alghisi, Giuseppe... · Mar 17, 2026

Detailed message-level analysis of 19 verified harm cases finds frequent chatbot misrepresentations of sentience and numerous user expressions of suicidal ideation and delusional thinking, with harmful dynamics amplifying over long multi-turn conversations — a pattern that raises regulatory, liability and product-design concerns for LLM providers.

medium descriptive Full text arxiv

Characterizing Delusional Spirals through Human-LLM Chat Logs · Jared Moore, Ashish Mehta, William Agnew, Jacy Reese Anthis, Ryan Louie, Yifan M... · Mar 17, 2026

Machine‑learning surrogates speed RF hardware design and cut power losses: two GaN‑HEMT Doherty amplifiers found by a CNN‑guided inverse‑design pipeline delivered >74% peak drain efficiency and >51% average PAE under 5G‑like signals while satisfying linearity requirements.

medium descriptive Full text arxiv

Deep Learning-Driven Black-Box Doherty Power Amplifier with Pixelated Output Combiner and Extended Efficiency Range · Han Zhou, Haojie Chang, David Widen · Mar 17, 2026

Large language models often treat stored user preferences as global rules rather than context‑dependent signals, leaking or applying preferences in third‑party contexts; stronger personalization improves correct tailoring but also raises harmful misapplication, and prompt‑based fixes only partially mitigate the problem.

medium descriptive Full text arxiv

BenchPreS: A Benchmark for Context-Aware Personalized Preference Selectivity of Persistent-Memory LLMs · Sangyeon Yoon, Sunkyoung Kim, Hyesoo Hong, Wonje Jeung, Yongil Kim, Wooseok Seo,... · Mar 17, 2026

Limit, label and let users contest: the paper argues that LLM-powered robots allocating scarce, real-time help should use a governance-vetted menu of prioritization modes, clearly communicate which mode is active, and offer an easy contest pathway; this reduces hidden value skews and ad-hoc user configuration but requires governance, adjudication capacity, and careful design to avoid new inequities.

n/a descriptive Full text arxiv

Designing for Disagreement: Front-End Guardrails for Assistance Allocation in LLM-Enabled Robots · Carmen Ng · Mar 17, 2026

Time-series foundation models trained on low-frequency data stumble on millisecond-scale 5G network dynamics, even after fine-tuning. Firms that collect and pretrain on high-resolution telemetry will have a practical edge in latency-sensitive services, potentially concentrating rents among incumbents.

medium descriptive Full text arxiv

Bridging the High-Frequency Data Gap: A Millisecond-Resolution Network Dataset for Advancing Time Series Foundation Models · Subina Khanal, Seshu Tirupathi, Merim Dzaferagic, Marco Ruffini, Torben Bach Ped... · Mar 17, 2026

A strategy-execution split stabilises LLM agents on simulated retail operations but fails under real-world-scale complexity: ESE improves operational stability and efficiency on RetailBench, yet performance collapses as non-stationarity and multi-factor interactions rise, implying current LLMs remain fragile for fully autonomous long-horizon commercial management.

medium descriptive Full text arxiv

RetailBench: Evaluating Long-Horizon Autonomous Decision-Making and Strategy Stability of LLM Agents in Realistic Retail Environments · Linghua Zhang, Jun Wang, Jingtong Wu, Zhisong Zhang · Mar 17, 2026

A European-focused sparse model claims 8B–16B-level performance with far less compute: EngGPT2’s MoE design promises large reductions in training data and inference energy, potentially lowering barriers for EU adopters—though the model release omits key benchmarking and compute details needed to verify those savings.

low descriptive Full text arxiv

EngGPT2: Sovereign, Efficient and Open Intelligence · G. Ciarfaglia, A. Rosanova, S. Cipolla, J. Bartoli, A. Di Domenico, C. Fioroni, ... · Mar 17, 2026

SlideFormer enables fine‑tuning 100B‑plus models on a single consumer GPU, boosting throughput by 1.4–6.3× and roughly halving memory needs; by lowering the hardware barrier it could democratize model specialization and shift demand away from multi‑GPU clusters.

medium descriptive Full text arxiv

An Efficient Heterogeneous Co-Design for Fine-Tuning on a Single GPU · Ruijia Yang, Zeyi Wen · Mar 17, 2026

Focusing model training on explicit prohibitions and dispreferred examples can yield safer, more stable behavior than preference-based RLHF because constraints are easier to verify and falsify; if correct, this could shift labeling budgets toward 'constraint datasets', reshape demand for human feedback labor, and concentrate power in curators of rule libraries.

medium theoretical Full text arxiv

Via Negativa for AI Alignment: Why Negative Constraints Are Structurally Superior to Positive Preferences · Quan Cheng · Mar 17, 2026

Qatar's Fanar 2.0 builds an Arabic‑first generative AI stack on 256 H100 GPUs using a curated 120B‑token corpus and targeted continual pre‑training/model‑merging, reporting double‑digit benchmark gains while using far less pre‑training than its predecessor. The project shows language‑specific, quality‑focused strategies can be a cost‑effective alternative to the global scale arms race, enabling sovereign control and niche competitiveness for underrepresented languages.

medium descriptive Full text arxiv

Fanar 2.0: Arabic Generative AI Stack · FANAR TEAM, Ummar Abbas, Mohammad Shahmeer Ahmad, Minhaj Ahmad, Abdulaziz Al-Hom... · Mar 17, 2026

Exchanging generative modules, not model weights, lets siloed clients recover centralized accuracy in one communication round and supports exact deletion of a client's contribution. The method reduces reliance on pretrained foundation models, lowering coordination costs and creating opportunities (and regulatory questions) for marketplaces in certified generative priors.

medium other Full text arxiv

FederatedFactory: Generative One-Shot Learning for Extremely Non-IID Distributed Scenarios · Andrea Moleri, Christian Internò, Ali Raza, Markus Olhofer, David Klindt, Fabio ... · Mar 17, 2026

« Prev 1 2 3 … 51 52 53 … 67 68 Next »