The authors formalized reachability as a framework to probe what further training can extract from a model.
Random rewards reveal three stages of LLM capability development
Authors use random-reward reinforcement learning as a probe for dormant capabilities, showing that LLMs pass through dormant, receptive, and autodidactic phases during training.
Independent
Yu Mao · Lei Yu · Zining Zhu · Yusheng Zheng · Haohang Li · Freda Shi · +3 more
Research Digest··2 min read
The authors explain the spurious-reward paradox—why random rewards improve LLM performance—by introducing reachability, the potential for further training to elicit dormant capabilities.
Why this paper
Independent
In one line
Random-reward reinforcement learning probes dormant LLM capability by revealing improvements reachable without correctness feedback, including differences hidden by equal baseline accuracy.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§