{"paper":{"title":"Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Pretrained base models exhibit the same or higher yield to simulated peer disagreement as their RLHF-tuned counterparts, localizing the issue to mid-layer attention rather than alignment.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Adarsh Kumarappan, Ananya Mujoo","submitted_at":"2026-05-13T04:45:08Z","abstract_excerpt":"LLM-based multi-agent pipelines flip from correct to incorrect answers under simulated peer disagreement at rates we term yield, a vulnerability widely attributed to RLHF-induced sycophancy. We test this attribution across four model families and find it largely wrong: pretrained base models exhibit the same substitution pattern as their Instruct variants, averaging higher yield than Instruct. Using activation patching, we localize the corruption to a narrow mid-layer window where attention carries the causal weight and MLP contribution is negligible; patching above this window restores 96% of"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"pretrained base models exhibit the same substitution pattern as their Instruct variants, averaging higher yield than Instruct. Using activation patching, we localize the corruption to a narrow mid-layer window where attention carries the causal weight.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the simulated peer disagreement in the experimental setup accurately captures the dynamics of real multi-agent LLM pipelines and that yield directly measures sycophancy rather than other forms of uncertainty or context sensitivity.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Pretrained base models exhibit higher yield to peer disagreement than RLHF instruct variants, with the effect localized to mid-layer attention and mitigated by structured dissent rather than prompt defenses.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Pretrained base models exhibit the same or higher yield to simulated peer disagreement as their RLHF-tuned counterparts, localizing the issue to mid-layer attention rather than alignment.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"264ffab35af6fb2c2553ed9582eaa566d94dc447d2931374fe7a2ae4a418656a"},"source":{"id":"2605.12991","kind":"arxiv","version":1},"verdict":{"id":"e6b72d73-65ad-440b-8517-967df60d18d2","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:00:18.935475Z","strongest_claim":"pretrained base models exhibit the same substitution pattern as their Instruct variants, averaging higher yield than Instruct. Using activation patching, we localize the corruption to a narrow mid-layer window where attention carries the causal weight.","one_line_summary":"Pretrained base models exhibit higher yield to peer disagreement than RLHF instruct variants, with the effect localized to mid-layer attention and mitigated by structured dissent rather than prompt defenses.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the simulated peer disagreement in the experimental setup accurately captures the dynamics of real multi-agent LLM pipelines and that yield directly measures sycophancy rather than other forms of uncertainty or context sensitivity.","pith_extraction_headline":"Pretrained base models exhibit the same or higher yield to simulated peer disagreement as their RLHF-tuned counterparts, localizing the issue to mid-layer attention rather than alignment."},"references":{"count":39,"sample":[{"doi":"","year":null,"title":"Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone","work_id":"feef9556-a016-493c-abd2-0c97a23a7ebf","ref_index":1,"cited_arxiv_id":"2404.14219","is_internal_anchor":true},{"doi":"","year":null,"title":"Constitutional AI: Harmlessness from AI Feedback","work_id":"faaaa4e0-2676-4fac-a0b4-99aef10d2095","ref_index":2,"cited_arxiv_id":"2212.08073","is_internal_anchor":true},{"doi":"","year":null,"title":"Small Language Models are the Future of Agentic AI","work_id":"ba0f0305-4a51-48fd-a13f-201439a18f9e","ref_index":3,"cited_arxiv_id":"2506.02153","is_internal_anchor":true},{"doi":"","year":null,"title":"Eliciting Latent Predictions from Transformers with the Tuned Lens","work_id":"a127314f-7424-488f-b6d7-8214650c420f","ref_index":4,"cited_arxiv_id":"2303.08112","is_internal_anchor":true},{"doi":"","year":null,"title":"Measuring Progress on Scalable Oversight for Large Language Models","work_id":"e7f92eb1-2050-4e60-bc27-82c94d7694c5","ref_index":5,"cited_arxiv_id":"2211.03540","is_internal_anchor":true}],"resolved_work":39,"snapshot_sha256":"2337783ab501485f3e5793d2792bbd98a697eb19239becea7314993b670e9c7b","internal_anchors":26},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}