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.
Researchers introduce Stateful CARS, an exact sampling method that reuses invalidity certificates across LLM agent execution histories.
Research shows model merging techniques degrade LLM safety alignment significantly more under adaptive jailbreaks than static tests indicate.
Research proposes the FCAC framework to help fraud ops safely authorize model automation under delayed audit labels and capacity limits.
Research demonstrates hybrid neural-classical correction techniques to adapt Google's frozen TimesFM model for high-frequency stock forecasting.
Research shows combining input transformations with output sampling improves LLM accuracy over standard self-consistency at matched compute.
Paper shows standard evaluation benchmarks for generative time-series models fail on sparse data with heavy zero or fixed-value point masses.
POLIS framework proposes institutional design rules—such as controlled delegation and resource bounds—to enforce multi-agent AI safety.
Research shows linear probes detecting LLM context corruption fail to accurately predict final answer correctness in multi-step reasoning.
RouteGuard proposes a theoretical framework to mathematically certify performance gain before deploying multi-agent LLM routing systems.
Researchers proposed spectral fingerprinting of weight space to trace open-weight LLM lineage and verify model provenance.
Researchers introduced BASIS, a method using prefill attention probes to selectively filter prompt injections without blanket model refusal.
arXiv study shows post-training quantization degradation varies heavily by task and scale, not just bit-width metrics.
arXiv research reveals LLMs' internal hidden states offer poorly calibrated signals for verifying generated code correctness.
Research shows safety degradation from LLM steering vectors comes from a separable component, enabling safer representation engineering.
Researchers propose an active-path auditing framework to evaluate privacy defenses and leakage channels in Retrieval-Augmented Generation.
Preprint demonstrates LLM miscalibration and severe overconfidence when generating outputs with missing critical context.
Researchers introduced TRACE, an automated feedback loop that mines agent execution logs to diagnose context and prompt errors.
Probing activations of open-weight reviewer models surfaces hidden code vulnerabilities missed by the model's textual outputs.
FinATOM introduces a head-free LLM interface predicting stock returns and ETF allocations directly via constrained token generation.
Researchers released BAD-ACTS, a benchmark and taxonomy for evaluating LLM agent robustness against adversarial attacks and harmful actions.
Researchers developed an automated pipeline to generate multi-turn conversational jailbreak attacks using psychological techniques.
New research demonstrates a Kernel Density Estimator framework to measure membership inference attack vulnerability in tabular synthetic data.
Research challenges synthetic tabular data benchmarks, demonstrating that evaluation metrics mask flaws in state-of-the-art diffusion models.
Research analyzes latent reasoning models, finding their non-text thought processes present severe explainability and monitoring challenges.
Research demonstrates a vision-based attack manipulating Computer Use Agents via visual attention concentration on GUI interfaces.
Study shows multi-agent yield under peer disagreement originates in base models, not RLHF, challenging standard consensus architectures.
Research demonstrates low-bit KV cache quantization causes silent safety alignment degradation across enterprise-scale open LLMs.
Researchers propose Wasserstein Distributionally Robust Regret Optimization to reduce over-conservatism in decision-making under uncertainty.
Researchers adapted Merlin-Arthur proof protocols to establish verifiable mutual-information bounds for RAG context accuracy.
Research paper clarifies Prediction-Powered Inference, a method enabling unbiased statistical estimates from imperfect ML proxy outputs.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion