Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
ArXiv paper addresses machine unlearning gaps by automating the identification of underlying source spans needed for data removal.
Researchers introduce DA-RAC, a distance-aware calibration method to fix context-induced miscalibration in LLM-as-a-judge evaluators.
ArXiv paper evaluates agent memory substrates, including dense, sparse, structured, and hierarchical stores for long-horizon tasks.
Researchers propose extracting temporal dependencies into structured graphs to run programmatic deterministic calculations for legal deadlines.
Researchers propose a novel method for identifying fine-grained hallucinated text spans and aligning them to input source evidence.
Researchers introduced MicroVerse, a framework measuring identity and goal drift in long-horizon multi-agent LLM simulations.
Researchers introduced PLSQLBench, a 2,865-instance execution-based benchmark evaluating LLM capabilities in procedural PL/SQL code generation.
KV-cache retention in LLM serving sessions causes stateful agents to attend to aborted transcript branches, breaking rollback consistency.
Research introduces R3-Bench showing LLMs fail to efficiently allocate shared computational budgets across multi-problem workloads.
ArXiv paper models hallucination compounding across multi-agent LLM pipelines, showing degradation of detectability at handoffs.
ArXiv study reveals recall-maximizing RAG retrievers degrade LLM code repair performance in fixed-budget contexts.
Latent structure analysis reveals standard human assessment instruments do not measure the same underlying constructs when applied to LLMs.
Research shows LLM verifiers become significantly more lenient when prior audit-repair history exists within the prompt context.
Research demonstrates that admission-time vector defenses fail against coordinated, low-signature document poisoning in RAG systems.
Audit reveals leading multi-hop RAG benchmarks rely on non-commercially licensed models like NV-Embed-v2 (CC-BY-NC-4.0), masking commercial IP risks.
Researchers introduced AWED-PIPER, an open-source framework with 54 expert detector models for PII anonymization across 36 languages.
A new empirical study reveals that compressing Chain-of-Thought reasoning traces to lower inference costs degrades model trustworthiness.
BiAxisBias framework reveals that 17.1% of LLM bias audit selections change under minor variations in prompt framing and context.
Researchers adapted the Honeyquest framework to evaluate how 21 different LLM architectures respond to cyber deception and honeypots.
Researchers introduced Forward-Pass-Only MLP training, enabling LLM domain adaptation with 3x throughput and 40% lower peak memory.
Empirical study on LexGLUE/FairLex shows combining Bayesian updating, Dempster-Shafer, and conformal prediction improves uncertainty calibration.
Research identifies statistical failure modes in field-level risk control for LLM document extraction, proposing valid calibration methods.
Researchers introduced DUET, an on-policy distillation method enforcing dynamic runtime prohibitions like PII redlines and policy rules.
Research introduces a framework to diagnose Causal Modality Sensitivity and perception-decision misalignment in multimodal LLMs.
WANDR introduces a 500-task benchmark evaluating autonomous research agents on wide entity discovery and deep, verifiable web investigation.
Research introduces a statistical audit framework showing net benchmark accuracy masks significant per-item shifts in compressed models.
Research introduces AgentRelBench, demonstrating that single-run evaluations miss catastrophic non-deterministic agent failure modes.
Researchers introduced SAGA, a generative action embedding model encoding user interaction sequences across multi-surface financial ecosystems.
Researchers introduced an post-hoc explanation method for StrGNN dynamic graph anomaly detectors used in fraud and intrusion detection.
Researchers introduced GraniKV, an asymmetric KV-cache paging layer that optimizes multi-agent inference by splitting prefix and suffix storage.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion