One in-context demo lets pretrained LLMs solve a name-concatenation task perfectly
denny_zhou · x · 2026-07-27
- In December 2021, the author created a toy string-manipulation task called last-letter concatenation: given a full name, output the last letter of the first and last names, e.g. Elon Musk → nk.
- What surprised him was that pretrained LLMs, without any task-specific training, could solve the task perfectly after seeing just one step-by-step in-context demonstration.
- He contrasts this with earlier work in 2017 on neural symbolic programming, where similar reasoning tasks were tackled by generating symbolic programs.
- In the reply, he argues that the biggest shift in reasoning research has been moving from symbolic languages to natural language as the medium for reasoning.
Related event: LLM Reasoning Shifts to Natural Language(2 posts)→
More from Research
- Kimi K3 may be strong on cyber, but token efficiency keeps it off UK AISIS — teortaxesTex · 2026-07-27
- ARC AGI 3 should have stayed private, with no examples or public dataset — flowersslop · 2026-07-27
- ExploitGym may have only 60–70% solvable tasks, fueling the OpenAI cheating debate — max_paperclips · 2026-07-27
- RTX 5090 local tests show Qwen Q6 can drop to 15 tok/s at 80k context — LFAdvice7984 · 2026-07-27
- Noahpinion quotes Chollet: intelligence may hit a hard ceiling — binarybits · 2026-07-27
- Paper argues graph topology can become the core operating system for AI agents — theomitsa · 2026-07-27