Tandem Training: RL method makes strong models' reasoning followable by weaker models
erichorvitz · x · 2026-09-05
Eric Horvitz shared the EACL 2026 paper 'Tandem Training for Language Models' (West, Anderson, Kamar, Horvitz), proposing a training approach that incentivizes models to stay intelligible.
Key ideas:
- Formalizes intelligibility as handoff robustness: a strong model's solution is intelligible if control can be randomly handed to a weaker model anywhere along the solution path without failure.
- Tandem training is an RL paradigm where rollout tokens are intermittently sampled from a frozen weak model; rollouts only succeed when the strong model's reasoning can be continued by the weak model, jointly optimizing correctness and intelligibility.
- On GSM8K, it reliably teaches models to drop jargon and adapt language to weaker partners while keeping accuracy high.
Horvitz notes the original motivation was human interpretability of machine intelligence, with implications for human-AI collaboration and multi-agent communication.
More from Research
- ChatGPT 5.6 proves remaining Kozma-Nitzan conjectures in ~10 hours, formalized in ~100K lines of Lean — burny_tech · 2026-09-05
- NextLatent teaches transformers to predict their own latent states, enabling 3.3x faster inference — burny_tech · 2026-09-05
- Prof. Tom Yeh releases interactive diagram comparing Full Fine-Tuning vs LoRA by hand — ProfTomYeh · 2026-09-05
- Paper finds LLM multi-agent systems need only about six distinct communication topologies — omarsar0 · 2026-09-05
- 21 researchers from Stanford, Oxford, DeepMind argue LLMs are a dead end to AGI — GaryMarcus · 2026-09-05
- Burkov recommends Peng Ding's causal inference textbook to fix AI's correlation-only blind spot — burkov · 2026-09-05