A single LLM can generate several parallel reasoning chains that adapt to each other at every token, improving coverage per token on enumeration, divide-and-conquer, and coding tasks.
Note that, as the newly generated token attends to all tokens in the KV cache, it also attends to the first token of agent 1
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Group Think: Multiple Concurrent Reasoning Agents Collaborating at Token Level Granularity
A single LLM can generate several parallel reasoning chains that adapt to each other at every token, improving coverage per token on enumeration, divide-and-conquer, and coding tasks.