Robust Context-Aware Detection of Malicious Instructions in Text
Researchers propose a text-segmentation defense methodology to detect indirect prompt injection (IPI) attacks in agentic LLM workflows.
arXiv cs.LG — Machine Learning·7 Aug 2026
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 propose a text-segmentation defense methodology to detect indirect prompt injection (IPI) attacks in agentic LLM workflows.
Researchers propose an unsupervised audit framework to detect and localize errors in autonomous data-analysis agents.
Researchers propose securing AI agent cryptographic operations by isolating private keys in hardware keystores rather than memory.
Researchers have developed finite-sample guarantees for localized conformal prediction to address subgroup-specific model miscalibration.
Researchers propose TS-RAG, a framework applying retrieval-augmented generation to time series forecasting by retrieving historical reference sequences.
Researchers introduce a scalable estimation framework for high-dimensional VARMA models, overcoming traditional computational limits.
Researchers propose ASAT, a method using human feedback to dynamically adjust thresholds for detecting out-of-distribution model inputs.
Researchers propose combining speculative decoding with online learning to adapt draft models dynamically, improving LLM inference speeds.
Researchers propose Fractal KV-Cache Archives to compress key-value cache symbol streams during long-context LLM inference.
A new study reveals that human reasoning inputs bias LLM outputs, causing models to adopt and amplify incorrect or harmful user assertions.
An academic study of 4,440 evaluation runs shows GraphRAG systematically underperforms vector RAG on citation precision for complex corpora.
Researchers propose scaffold-mediated post-training, co-evolving LLM parameters and procedural scaffold graphs to internalize complex strategies.
Researchers propose "circuit anchors" to prevent self-evolving language models from losing core safety and alignment capabilities during optimization.
Researchers audited privacy leakage in multilingual RAG pipelines, finding vulnerability patterns across stages using Qwen2.5-7B.
Researchers demonstrate cross-model steering transfer, showing concept directions from one LLM can control behavior in a different model.
A research paper demonstrates that LLMs adopt cognitive biases based on user interaction history, even when instructions demand objectivity.
Researchers introduce SkillZip, a graph compression method for LLM agent skill libraries to reduce context window and inference costs.
Researchers propose FOCUS, a method to decouple LLM expert personas to prevent cross-domain behavioral leaks in risk-sensitive domains.
Researchers propose using random number generation to mitigate context-insensitive scoring bias in LLM-as-a-judge evaluation frameworks.
Researchers propose a zero-shot LLM detection method via latent prompt restoration to identify machine-generated text.
An arXiv research paper evaluates speech LLMs against context biasing methods for recognizing rare and domain-specific words in speech recognition.
Researchers introduce HallDetect, a reference-free, black-box framework for detecting LLM hallucinations in source-grounded tasks.
Academic researchers introduced Synthetic Query Probing, a method to map similarity score distributions across different embedding models.
Researchers propose MACRO, a Markov chain routing method for transformer layers to dynamically skip or repeat layers without retraining.
Researchers prove that routing web agent tasks to the optimal visual/text observation mode is mathematically difficult to learn.
Researchers benchmark LLM capabilities in reviewing complex, rule-intensive national standard documents to improve consistency and compliance.
Researchers introduced a reference-free framework using LLM judges to evaluate the quality, consistency, and complexity of conversational agent benchmarks.
Academic research evaluates programmatic code execution against structured JSON tool-calling, highlighting trade-offs in agent capabilities.
Researchers introduce MIST, a benchmark evaluating an LLM's ability to selectively trust external context over internal knowledge.
An academic paper introduces Agentic Nesting, a methodology to orchestrate and integrate fragmented legacy enterprise applications using LLMs.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion