Research Feed

The paper introduces a dual-domain architectural pattern that separates an AI agent's persona from its execution logic to improve governance and auditability in regulated environments.

Researchers evaluated how medical Large Language Models reason by systematically perturbing inputs and measuring if the model's logical chain stays consistent with its final output.

The paper demonstrates that standard LLM agent handoff processes often cause binding constraints to lose their functional power, and identifies methods to restore this operational state.

The paper introduces a method that models agent behavior as finite state machines to improve next-step prediction and detect system failures.

The paper demonstrates that simple linear probes on frozen model hidden states provide efficient and robust detection of machine-generated text using minimal training samples.

The paper demonstrates that language models can be steered during inference using undisclosed logit modifications, making traditional model weight audits insufficient for identifying production-level bias.

The paper introduces a gated intervention framework that dynamically manages model activations to reduce sycophancy and hallucinations in clinical question answering while preserving model weight integrity.

Semantic Overlays add an out-of-band annotation layer to input streams to help models distinguish between trusted instructions and untrusted content, significantly reducing prompt injection success rates.

StepGuard provides a 4B guard model that validates agent tool actions before execution to reduce unsafe behaviors while maintaining performance.

The researchers introduce a penalty method to prevent language models from sacrificing safety protocols when they are fine-tuned for improved reasoning tasks.

The paper introduces a structural audit procedure that reliably detects and localizes causal leakage in complex sequence models by monitoring intermediate output differences during forward passes.

Researchers developed a benchmark called FORGE to measure how easily LLMs can be tricked into recommending fake products through search-augmented content.

Researchers developed an injection attack method called InjecMEM that can override an agent's memory by manipulating the content stored in its retrieval systems.

SecOPD improves AI agent security against adaptive prompt injection by using on-policy distillation to provide fine-grained training signals that distinguish between trusted instructions and malicious data.

Researchers tested whether language models can detect and report on internal computational interventions, finding that model confidence signals are more informative than direct verbal reports.

ReFrame acts as a secure intermediary layer that analyzes and rewrites potentially unsafe multimodal inputs before they reach closed source models.

The paper introduces a method to calculate the effective sample size for thresholding models when calibration data is clustered and contains correlated errors.

The paper introduces a rank-calibrated detector called TRACE-C designed to identify anomalies in complex electricity system telemetry.

The paper introduces a method to improve privacy policy adherence in LLMs by intervening on specific attention heads to align model outputs with user-defined privacy preferences.

AID-Guard ensures that AI agent decisions result in exactly one provider effect by binding user intent to durable, stateful authorization protocols.