LetTokenBurn
Sign in
Auto86%Harbor: pendingSafety & AlignmentInstruction Following

RobustReview

A Missing Piece for Trustworthy AI Reviewers: From Benchmarking Rhetorical Robustness to SciCore Review

A controlled full-manuscript benchmark with 1,260 versions evaluating LLM reviewers on rhetorical robustness (stability across rewrites vs discrimination between papers).

PaperRepo

How to run

ltb · robustreview
$ npm i -g lettokenburn-cli
$ ltb bench info robustreview

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

8 sourced entries
HarnessSourceConfidenceLink
GPT-5.50.6N/AIndependentmediumSep 30, 2026arxiv.org
Claude Sonnet 50.6N/AIndependentmediumSep 30, 2026arxiv.org
GPT-5-mini0.6N/AIndependentmediumSep 30, 2026arxiv.org
Qwen3.5 Flash0.6N/AIndependentmediumSep 30, 2026arxiv.org
GLM-5.20.6N/AIndependentmediumSep 30, 2026arxiv.org
Kimi K2.60.6N/AIndependentmediumSep 30, 2026arxiv.org
GPT-oss 120B0.5N/AIndependentmediumSep 30, 2026arxiv.org
Gemini 3.5 Flash Lite0.5N/AIndependentmediumSep 30, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

Construction: Built from 60 anonymized ICLR 2026 submissions. Each paper is rewritten under 10 rhetorical conditions by two independent LLM systems to create variants preserving scientific content but changing presentation.

Scoring & Findings: Evaluates reviewer configurations on joint stability-discrimination and human alignment. Reveals 'false robustness' where low rewrite sensitivity correlates with score collapse, showing that existing evaluations often fail to distinguish between rhetorical manipulation resistance and lack of discrimination.