Asai et al. (2023) — Self-RAG, and whether a model can grade its own retrieval

[akash-coded]

Reading Club. One paper per thread. The house position: replicate, do not cite.
A paper we can re-run against this corpus is worth far more argued about with numbers.

Paper: arXiv:2310.11511

The claim, in one paragraph

The model emits reflection tokens deciding whether to retrieve, and whether what came back
supports the answer. Retrieval becomes conditional and self-assessed rather than unconditional.

The interesting claim is not the architecture. It is that a model can usefully judge the
sufficiency of its own evidence — which is precisely the thing this repository has failed to do
with a threshold.

How we would test it here

Issue #10 is still open: no retrieval-score threshold separates answerable
from unanswerable, best F1 0.38 at the real base rate. #35 argues the
sufficiency check should be a model call rather than a classifier, and lands on a cascade.
nanorag/agent.py has the sufficiency hook to plug into.

Two or three things to argue about

1. Self-assessment has an obvious failure mode: confident agreement. The model that produced the answer is grading the evidence for the answer. Is the reflection token independent enough to be worth anything?
2. We measured four surface signals near chance. Self-RAG's claim is that entailment needs a reader. Does that make our negative result a finding about thresholds, or about our corpus?
3. Cost. Reflection tokens on every query. At what abstention-error cost does that become obviously worth it, and can you write that as a number rather than a preference?

---

If you have run any part of this, post the numbers — a replication with an interval outranks
every opinion in the thread, including mine.

[GioiaZheng]

I would separate two questions that Self-RAG tends to blur together:

1. did retrieval return enough evidence to answer?
2. does the generated answer agree with that evidence?

The first one should be measured before generation if possible. If the same model first writes an answer and then grades whether the evidence supports it, the failure mode is not just overconfidence; it is answer-conditioned agreement. The grader is no longer judging the evidence set, it is judging evidence through the shape of the answer it just committed to.

A small experiment I would run here:

condition A: grader sees question + retrieved nodes only
condition B: grader sees question + retrieved nodes + generated answer
condition C: separate small reader/grader sees question + retrieved nodes only

Then evaluate all three on four slices:

answerable_in_top_k
answerable_in_corpus_but_missing_from_top_k
not_answerable_in_corpus
conflicting_or_partially_supporting_evidence

The interesting metric is not just F1. I would report false-positive rate on the two negative slices, because that is where a RAG system creates unsupported answers. A model that slightly improves answerable recall but lets unsupported cases through is usually worse in production.

For the cost question, I would write the decision rule as expected cost rather than preference:

use_reflection_if:
  C_bad_answer * (fp_without_reflection - fp_with_reflection)
  + C_unneeded_retrieval * retrievals_avoided
  > C_reflection_call
  + C_missed_answer * (fn_with_reflection - fn_without_reflection)

In most support / documentation RAG systems, C_bad_answer is much larger than C_missed_answer, so a reflection step can be worth it even if it increases abstentions. In exploratory search, the costs may flip.

So my prior is: the negative result from fixed thresholds is probably a finding about surface signals, not necessarily about the corpus. But I would only trust Self-RAG-style reflection if it is calibrated on unsupported-evidence slices and if the sufficiency judgment is made before the answer is generated.

I have not run this notebook yet, so this is a proposed test rather than a measured result.