345K trajectories reveal LLM agent personalities: Claude most planful, GPT-5.5 least energetic
mdredze · x · 2026-10-23
The paper "On the Behavioral Traits of LLM Agents" (arXiv:2609.32776) proposes A-B-D: inferring agent traits bottom-up from behavior instead of self-reports or LLM judges.
- 345,667 real-world trajectories spanning 80 models, 12 tasks, and 50 harnesses, yielding 318 candidate functional and linguistic features.
- After filtering for stability and discriminability, 79 features factor into six stable, model-attributable factors (two functional, four linguistic).
- Findings: Claude-Sonnet-4.6 and Kimi-K3 show the most planfulness; GPT-5.5 and GPT-5.6 are the least energetic.
- The "knowledge-action gap" is quantified: behavioral factors correlate only weakly with self-reported Big Five scores (e.g., extraversion vs energetic, r = 0.07, p = 0.58) — what models say about themselves diverges from what they do.
Related event: Study Infers Personality Traits of 80 LLM Agents From 345K Trajectories(2 posts)→
More from Models
- Smaug-Nemotron-3.5 debuts as a top US open-source agentic coding LLM that runs on a single machine — bindureddy · 2026-10-23
- DeepSeek on RDNA2/3 + DDR4: 19 t/s peaks, 9 t/s average decode on 56GB VRAM — EmPips · 2026-10-23
- IBM CEO: best open AI models are Chinese, West 6-12 months behind, 75-90% of tokens going open-weight — victor_explore · 2026-10-23
- ArchitectureIQ Benchmarks LLMs' Training Intuition: GPT-5.6 Sol Hits 76.4% vs 33% Chance — ZimingLiu11 · 2026-10-23
- Abliterated GLM-5.3 sells for $10 as "world-class penetration tester", fueling vulnerability surge — deedydas · 2026-10-23
- DeepSeek harness slashes tokens, time and cost vs Codex, but tanks on FrontierCode — teortaxesTex · 2026-10-23