Skip to content

Reviewer Independence: Context Reset or Model Diversity

Reviewer independence comes from a fresh context, a different model family, or new evidence, each addressing different errors (inference).

A reviewer can differ from the author in the conversation it sees, the model it runs, and the evidence it reads. When a Second Model Helps in LLM Review holds the measurements that compare the first two. The table below maps them to error classes.

Two pages on this site take opposite sides. The critic agent pattern asks for a different model. The Elephant-Goldfish Model (EGM) runs its Claude reviewers with model: inherit, so reviewers use the starting session's model (EGM Claude integration).

What a fresh context removes and keeps

EGM's design text says "Self-review by the authoring LLM suffers from severe confirmation bias" (review.md). Its fix is a new reviewer that gets no parent deliberation (Claude integration). Reviewers still read files, and the zoom-out reviewer must search the codebase for reuse (zoom_out).

The reset keeps the weights. Song's fresh-session condition also withholds the original request, so it cannot isolate context reset (v2, §3.2). Its findings overlap same-context review at a rate comparable to within-model session pairs (§5.3). The study uses Korean artifacts (limitations).

The EGM page covers Self-Correction Bench. Tsui scopes the claims "to conversational role rather than authorship" (Tsui, §4.2). Reasoning models show "a small, even negative" blind spot (§4.4).

What a different model family adds and keeps

In Song's data, cross-model pairs shared fewer findings than sessions of one model (§5.3). One fresh-session review plus one cross-model review also matched more planted errors than two fresh-session reviews. That gain was not significant against two reviews by the stronger cross-model reviewer, so capability and model difference stay confounded (§5.2).

A second family can still share errors (inference). Across 9 judges from 7 families, "the models make the same mistakes on the same items" (Kohli). On MNLI, same-family pairs were 0.047 more correlated than the cross-family mean (§5.3). That study scored natural language inference labels, not review.

Chen et al. measure a model "favoring its exact own output", not its family (limitations). They also find that a long chain of thought before evaluation reduces harmful self-preference (abstract).

Adding a second family requires another review call and integration (inference). Claude Code's subagent model field documents Claude aliases, Claude model IDs, and inherit (Claude Code subagents). A cross-family reviewer there needs another host or a gateway (inference).

Evidence the author never read

Ao, Gao, and Simchi-Levi note that "Planner, worker, critic, and reviewer modules are often built from the same model family" (arXiv 2603.26993v1, §1). Their Table 1 counts a shared base model and a common prompt as a common signal. Tool calls and API outputs are new private signals (Table 1). Corollary 9 says rereading the terminal information state cannot improve Bayes risk, while nonredundant verifier signals can (Corollary 9). The paper does not test reviewers from different families.

Under that frame, a shared rubric adds no new signal (inference). A code read may add one if it supplies nonredundant evidence (inference). EGM passes a document standard to zoom-out (egm-review). The comprehension reviewer gets "no document standard" (egm-review). The codebase_details reviewer checks the document against source code (persona).

Decision table

Cells are inference unless marked observed or measured. Song's category F1 estimates lack uncertainty intervals and significance tests (§5.4).

Condition Same model, fresh context Add a different family Source
The author's conversation hid a gap Targets this No specific gain review.md
A test, tool, or code read can check the claim Give the reviewer that access No specific gain Ao et al., Corollary 9
Code in Song's study Higher observed F1 Lower observed F1 Song, Table 3
Documents in Song's study Lower observed F1 Higher observed F1 Song, Table 3
The judge scores its own exact output Reason at length before judging Removes the own-output case Chen et al.
Many reviewers will vote Correlated Still correlated (measured) Kohli

Key Takeaways

  • Name the error class first; choose context, model, or evidence changes to test against it (inference).
  • Song changes context and request access together; use evidence checks to test gaps a reset could leave (inference).
  • Allow for correlated errors when combining reviewers from different families (inference).
  • Do not cite Self-Correction Bench for a second model; it tests conversational role inside one model.
Emerging — single recent source Last reviewed: 2026-11-03