The authors trained a soft prompt and layer-weighting parameters with a contrastive objective, while leaving the underlying language model unchanged.
One LLM mechanism handles both retrieval and long-context evidence selection
UNREAL uses a frozen language model’s internal representations to rank evidence across prompts and a 21-million-chunk Wikipedia index.
Big Tech
Edan Kinderman · Elad Hoffer · Yochai Blau · Brian Chmiel · Ron Banner · Daniel Soudry · +1 more
NVIDIA · Technion
Research Digest··3 min read
Kinderman et al.
Why this paper
From NVIDIA and Technion
In one line
A single frozen LLM with fewer than 500K trainable parameters handles both corpus retrieval and long-context evidence selection.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (5 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§