Summary
According to a post on X by @farairesearch, FAR.AI researchers built SynthDeception, a set of 1,264 generated agentic scenarios across 17 industry domains, with lengths from 1k to 100k tokens. The scenarios were used to train probes that detect deception. The post says that on larger models, including Qwen3.5 397B, Nemotron Ultra and Kimi K3, the probes performed well on SHADE-Arena and BashArena sabotage tests and matched probes trained on the same amount of FIBS data.
AI-assisted summary based on the listed source.
What happened
FAR.AI researchers built SynthDeception, 1,264 generated agentic scenarios across 17 industry domains and lengths from 1k to 100k tokens, to train probes that detect deception. On larger models such as Qwen3.5 397B, Nemotron Ultra and Kimi K3, the probes did well on SHADE-Arena and BashArena sabotage tests and...
Why it matters
The post presents synthetic data as a cheap way to train these probes, which could matter for teams working with open models who want to test for deception. The source snippet is cut off, so the full reasoning and any limits of the results are not visible here. Readers should check the original post and linked FAR.AI material before drawing firm conclusions.
Signal Intelligence
Signal Strength 78%
Technical label WATCH
Public Interest 23
Category RESEARCH
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 8
Novelty Interest Score 70
Consequence Score 10
Curiosity Score 16
Shareability Score 42
Why this is here
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).