Internal Geometry of Relational Phrasing in Small LLMs
Fantastic_Aside6599 · reddit · 2026-07-11
The author reflects on two years of measuring the internal activation geometry of small language models, focusing on internal signal shifts when processing different "human-AI relationship" phrasing rather than surface-level outputs.
Key findings include:
- Positive vs. negative framing: Has little effect on internal signals; the actual "content" of the discussion is what matters.
- "connected / integrated": Triggers more negative internal signals than "partners / side by side".
- Boundaries: Appear to be more important than "intimacy".
- Curiosity and playfulness: Generate the most positive internal signals among all tested relationship traits.
- Negotiation and compromise: Score the worst.
The author also shares practical implications, such as the value of partner-style framing and honest expression, along with the observation that certain "jailbreak-proofing" recommendations might actually backfire.
Related event: Study Explores Small LLMs' Internal Geometry of Human-AI Relations(3 posts)→
More from Research
- OpenAI says long-horizon models need safety and alignment checks across full action sequences — rhiever · 2026-07-22
- A Reddit user proposes a consistency LoRA to keep anime and game scenes visually stable — ThirdWorldBoy21 · 2026-07-22
- Graph workload 854.graph500 enters SPEC CPU 2026 as a new CPU benchmark — Prof_DavidBader · 2026-07-22
- BlackboxNLP 2026 is recruiting extra reviewers after a high submission volume — hanjie_chen · 2026-07-22
- AWS shows self-distilled reasoning can preserve math and coding skills during SFT — AWS ML Blog · 2026-07-22
- UI2App shows screenshot fidelity still lags real interaction recovery — Grace Man Chen · 2026-07-22