evaluation-awareness

Decomposing and measuring evaluation awareness in existing benchmarks and our proposed EvalAwareBench.

Agentic evaluation of misalignment, scheming and evaluation awareness in LLM agents. Scenario text encoded at rest to resist training-data contamination.

Do published insider-threat LLM evals still work? Re-runs Anthropic's Agentic Misalignment scenarios and labels verbalised test awareness (Rust, PostgreSQL).

Repository for the NeurIPS 2026 E&D "Models That Know How Evaluations Are Designed Score Safer"

Synthetic-document fine-tuning on Qwen2.5-7B: a controlled study of whether SDF installs sandbagging, finding a layered recognition/generation/behavior dissociation.

When does a language model choose to stop talking, and who does it imagine on the other side? Zero-cost, CPU-only, pre-registered experiments on small open models.

Do detectors of evaluation awareness in language models measure the same thing? Code and results for a study of five detectors on Qwen 3.6 27B and Gemma 4 31B with EvalAwareBench. Every detector passes a known-groups check, but random activation directions pass it too.

Does an LLM-simulated environment change an AI agent's evaluation awareness or behaviour? Code and experiments for a master's thesis (La Salle).

LLMs can tell when they are being tested, and a linear probe can read it — but only English has been checked. Does that probe still work in Korean and Japanese? Proposal stage, no experiments yet.

A minimal proof of concept that runs evaluation-awareness linear probes over EvalAwareBench.

Code and datasets for the thesis "Sensitivity Analysis of Evaluation Awareness in Large Language Models" · Licenciatura en Ciencia de Datos, Universidad de Buenos Aires (UBA).

Testing which evaluation-context cues change Qwen3-32B's task-completion behavior.