Preprints and papers, rewritten in plain English.
The paper reframes long-video agent memory as a context-management problem, proposing a meta-optimization harness to fit hours of multimodal evidence into bounded token and latency budgets.
A new method, HoloCosm, selects benchmark samples by considering multiple evaluation metrics at once, potentially lowering cost while preserving multi-dimensional coverage.
A new arXiv paper proposes learned token pruning during prefill, mirroring how human readers skim large inputs.
A new method reuses rejected verifier states to overcome the fixed-length block limitation in block diffusion drafters.
A new arXiv paper proposes latent multimodal memory reasoning to help personalized agents handle distributed history across modalities and evolving dialogue states.
New research examines how the geographic context of a query affects which sources and brands appear in generative search answers.
A new data-collection framework targets the real bottleneck in low-resource speech recognition: scarce transcribed spontaneous speech.
New research examines how hidden malicious behaviors in foundation models can emerge after fine-tuning and proposes defenses for language and multimodal agents.
A new study compares base, fine-tuned, and preference-optimized models across seven lineages to see how post-training shapes cooperation in repeated interactions.
A new arXiv paper proposes automatically mutating detection rules to cover attack variants that evade original signatures.
A new arXiv preprint suggests using heart-signal vibrations as a biometric that TinyML devices could verify.
A new end-to-end approach first locates patch-relevant code, then reasons about it, avoiding brittle function-level matching.
A new decision layer for vehicular ad hoc networks abstains from binary validation when evidence is insufficient, targeting the understudied step between trust scores and driver action.
A new integrity-checking method lets agent networks verify data without false mismatches caused by serialization differences.
A new paper offers a systematic sensitivity characterization as an alternative to the Jackpot policy in Pearl blockchain's proof-of-useful-work scheme.
A new arXiv preprint investigates whether task-specific security skills improve LLM coding agents' ability to find vulnerabilities in the Linux kernel.
A new arXiv paper outlines how wirelessly captured evidence could help determine whether spoofed or jammed GNSS and AIS signals caused a maritime accident.
A new arXiv paper targets alert flooding, a major obstacle to practical deployment of provenance-based intrusion detection systems against APTs.