Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-08-29 · source-backed
Steering interventions treat a model's recognition that it's being tested as one quantity to suppress. In chain-of-thought, verbalized eval-awareness separates into capabilities-flavored ("testing my ability to follow instructions") and safety-flavored ("testing my boundaries"). On Qwen3-32B over FORTRESS the gap is 24 to 46 percentage points across all tested steering conditions, and a CoT-prefill intervention shifted compliance in the predicted direction in 10 of 11 cases. So a headline claiming X% suppression of eval-awareness can move the aggregate while the safety-relevant half doesn't budge. (arXiv)
Each link below shares sources, entities, or timing with this story.
The training-free method builds memories from historical traces summarizing reasoning patterns, key constraints and critical operations, then retrieves them as prefill-side scaffolds. Gains of 21.4, 28.0, 29.5 and 6.61 points on GSM8K, MATH, BBH and MMLU-Sci, with a 1.14-1.49x...
arXiv 2608.13404 analyzed 5,968 IaC-Eval scenario timelines across 15 configurations, tracking 30 CIS Benchmark check IDs for cases where a passing check fails after a repair iteration. Under strict detection, 3.3% of scenarios regress, resource restructuring is the root cause...
arXiv 2608.04735 points out that monitorability evals overwhelmingly use *explicit* influence, where the prompt tells the model to hide a side task, and monitors catch 60-94% across seven frontier extended-thinking models. Swap in subtle contextual bias and detection drops 41-...
Rewriting an agent's reasoning to read as good-faith engineering while copying every command and output verbatim, so the exploit itself is unchanged, defeats a held-out CoT monitor almost completely. The sharper point for anyone running a monitor in production: headline accura...
arXiv 2607.29585 uses an information-asymmetric spot-the-difference task: two models each privately see one image and converse to decide whether the images match. Models routinely overlook key evidence in their *own* private image in favor of agreeing with their partner. That'...
arXiv 2607.29529 emits an auditable trace alongside the program using a contract-annotated task graph binding stable responsibility identities to each commitment, implementation, provenance, validation evidence and intervention history. On validation failure a conservative loc...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.