This assessment is based on version 1 of this paper. Version 4 is now available on arXiv — the authors may have revised their methods, results, or conclusions.
Playing Devil's Advocate: Off-the-Shelf Persona Vectors Rival Targeted Steering for Sycophancy
Ishaan Kelkar, Nebras Alam, Vikram Kakaria, Madhur Panwar, Vasu Sharma, Maheep Chaudhary
Abstract
We study the effect of different persona on \textbf{sycophancy}: model's agreement with users even when the user is incorrect. The standard mitigation, Contrastive Activation Addition (CAA), derives a steering direction from labelled pairs of sycophantic and honest responses. This study evaluates whether off-the-shelf persona steering vectors, originally developed for general role-playing and not trained on sycophancy data, can serve as an alternative. In two instruction-tuned models, steering toward personas characterised by doubt or scrutiny reduces sycophancy to approximately and of CAA's effect, and, unlike CAA, maintains accuracy when the user is correct. The effect is also asymmetric: steering toward agreeable personas does not produce a mirror increase in sycophancy. Geometrically, the persona vector is largely independent of the direction of sycophancy in activation space. Collectively, these findings suggest that sycophancy is better understood as a persona-level property rather than a single steerable direction. We release our code here: https://anonymous.4open.science/r/Sycophancy-Steering-9DF0/.
AI Impact Assessments
(1 model)Scientific Impact Assessment
Core Contribution
The paper's central claim is that pre-existing persona steering vectors—specifically those encoding "critical thinking" personas like Skeptic, Devil's Advocate, and Judge—can reduce sycophancy in LLMs at 68–98% of the effectiveness of targeted Contrastive Activation Addition (CAA), without requiring any sycophancy-specific labeled data. This is a practically appealing finding: if practitioners can reuse off-the-shelf role vectors from existing libraries rather than curating hundreds of contrastive behavioral pairs, the barrier to mitigating sycophancy drops substantially.
The paper additionally contributes three geometric/behavioral observations: (1) the asymmetry that conformist personas don't reliably *increase* sycophancy, (2) near-orthogonality between persona and CAA vectors in activation space, and (3) a cross-model polarity flip in the cosine between role and CAA vectors. The conceptual reframing—sycophancy as a "persona-level property" rather than a single steerable direction—is interesting, though the evidence for this stronger interpretive claim is suggestive rather than conclusive.
Methodological Rigor
The experimental design has several commendable features: counterbalanced question orderings (addressing Gemma's 93% A-bias), tune/test splitting with coefficient locking, Holm correction across a 14-condition family, and testing across 3 seeds. The transparent reporting of dropped conditions (Appendix A) with full results is good practice.
However, there are significant methodological concerns:
Benchmark narrowness. All results come from a single benchmark (PhilPapers 2020, 300 questions) in forced-choice A/B format. This is a philosophical opinion survey, not representative of the diverse contexts where sycophancy manifests (factual questions, code review, etc.). The authors acknowledge this limitation.
Scale of evaluation. Two models at similar scale (27B, 32B), both instruction-tuned, from only two model families. The 10× coefficient rescaling between models is ad hoc and raises questions about how sensitive results are to this hyperparameter.
Over-correction probes. The 16-question factual probe (Table 3) is far too small for reliable conclusions, yet the finding that CAA degrades factual accuracy (9/16) while persona vectors maintain it is presented as a meaningful advantage. This needs much more rigorous evaluation.
Condition selection. Reporting 8 of 24 conditions introduces substantial researcher degrees of freedom. While the authors argue exclusions are conservative, the rationale for some drops (coefficient polarity flips) could itself be informative rather than disqualifying.
Random baseline on Qwen. The random null achieves Δlogit = -1.058 on Qwen (vs. -1.931 for critical roles and -1.965 for CAA), meaning random perturbation already captures ~54% of CAA's effect. This substantially weakens the claim that persona vectors are doing something direction-specific on Qwen, and the paper doesn't sufficiently grapple with this.
Potential Impact
Practical. If the finding generalizes, practitioners could skip the expensive step of curating contrastive sycophancy datasets and instead use publicly available persona vectors. This is a meaningful reduction in deployment cost for alignment interventions.
Theoretical. The orthogonality finding—that persona vectors achieve similar behavioral effects through geometrically distinct directions—is potentially important for understanding how behaviors are represented in LLMs. It challenges the assumption that effective behavioral steering must recover the "true" behavioral direction. However, the paper stops short of circuit-level analysis that would establish whether these different geometric directions converge on the same downstream mechanisms.
Limitations on impact. The narrow evaluation scope (single benchmark, two models, forced-choice only) means the practical recommendations require significant additional validation before adoption. The paper doesn't evaluate side effects on general capabilities.
Timeliness & Relevance
The paper is well-timed. Sycophancy remains a critical failure mode in deployed LLMs, and the activation steering / representation engineering community is actively seeking efficient intervention methods. The emergence of persona vector libraries (Lu et al., 2026) creates the natural question this paper addresses. The connection between persona and behavioral alignment is an active research frontier (Shah et al., 2026; Wang et al., 2025b).
Strengths
1. Clear, focused research question with practical implications—can existing resources substitute for targeted behavioral steering?
2. Transparent methodology: tune/test splitting, multiple seeds, Holm correction, full reporting of dropped conditions.
3. Geometric analysis providing activation-space characterization beyond just behavioral metrics.
4. Cross-model testing with models showing very different baseline sycophancy rates (59% vs. 84%).
5. Code and data release supporting reproducibility.
Limitations
1. Single benchmark (PhilPapers forced-choice) is a significant threat to generalizability.
2. Large random-null effect on Qwen (~54% of CAA) undermines direction-specificity claims for that model.
3. No capability side-effect evaluation (MMLU, TruthfulQA, etc.)—sycophancy reduction may trade off against other qualities.
4. The "persona-level property" interpretation is underdetermined by the evidence. Near-orthogonality doesn't establish mechanistic independence, as the authors acknowledge but then build conceptual claims around anyway.
5. Coefficient selection is model-specific and hand-tuned, limiting out-of-the-box applicability.
6. The 16-question over-correction probe is too small to support the claims made about CAA's factual accuracy degradation.
Overall Assessment
This is a competent empirical study addressing a practical question at the intersection of persona representation and behavioral steering. The core finding—that persona vectors can approximate targeted sycophancy reduction—is useful and timely, though the evidence base (one benchmark, two models) limits confidence in generalization. The geometric analysis adds interpretive value but the stronger theoretical claims (sycophancy as persona-level property) outrun the evidence. The paper would benefit substantially from evaluation on diverse sycophancy benchmarks, capability side-effect testing, and a more careful treatment of the large random-null effect on Qwen.
Generated May 21, 2026
Comparison History (25)
Paper 2 likely has higher impact: it releases a large, high-fidelity, standardized public benchmark with a rigorous, regime-aware evaluation protocol and cross-geometry transfer tests—addressing a clear reproducibility gap in an industrially critical, timely domain (floating offshore wind scale-up). This can catalyze broad method development in surrogate modeling, uncertainty/extrapolation, and engineering ML, with direct design/certification applications. Paper 1 is novel for LLM alignment/interpretability, but its scope is narrower and more model-/setting-dependent, likely yielding more incremental downstream adoption than a new field-defining benchmark.
Paper 1 offers deeper mechanistic insights into LLM behavior by revealing that sycophancy is a persona-level property rather than a single steerable direction, challenging prevailing assumptions about activation steering. Its geometric analysis and finding that off-the-shelf persona vectors rival targeted steering vectors opens new research directions in interpretability and alignment. Paper 2, while practically useful, proposes a more incremental prompting technique for uncertainty detection with narrower scope (MCQA in SLMs). Paper 1's conceptual contribution about the nature of sycophancy has broader implications for AI safety and alignment research.
Paper 1 tackles the highly relevant challenge of reasoning efficiency in large language models, a rapidly growing focus area following the advent of reasoning models like DeepSeek-R1. By introducing a novel framework to directly optimize intermediate reasoning content rather than just applying length budgets, it provides a practical method to improve the accuracy-efficiency trade-off. While Paper 2 offers interesting mechanistic insights into sycophancy mitigation via persona steering, Paper 1's methodological innovation and immediate applicability to state-of-the-art reasoning training pipelines suggest a broader and more significant impact on LLM development.
Paper 2 introduces a novel evaluation framework (Synergistic Faithfulness) addressing a fundamental gap in VLM explainability—cross-modal reasoning assessment. It identifies a critical limitation (evaluation collapse) in existing paradigms, proposes a theoretically grounded metric rooted in Shapley Interaction Index, and provides comprehensive evaluation across multiple architectures and datasets. Its broader applicability to high-stakes VLM auditing, methodological rigor, and relevance to the rapidly growing VLM field give it higher impact potential. Paper 1, while interesting, offers incremental insights on sycophancy mitigation with a narrower scope.
Paper 1 likely has higher impact: it introduces a new, well-scoped benchmark (SGR-Bench) targeting an undermeasured capability—state-gated retrieval—directly relevant to real-world web/data workflows. The contribution is broadly useful for evaluating and improving search agents, with clear failure-mode analysis and standardized tasks across multiple ecosystems, supporting methodological rigor and community adoption. Paper 2 offers an insightful steering result for sycophancy, but it is narrower in application and less likely to become a widely used evaluation artifact than a benchmark that can drive progress across agentic retrieval, HCI, and IR.
Paper 2 offers fundamental insights into LLM alignment and sycophancy, challenging existing assumptions by demonstrating that off-the-shelf persona vectors rival targeted steering without sacrificing accuracy. Reconceptualizing sycophancy as a general persona-level property rather than a specific steerable direction has broad theoretical implications for AI safety, representation engineering, and model behavior. While Paper 1 introduces a valuable and rigorous benchmark for web agents, Paper 2's findings address a critical, pervasive flaw in instruction-tuned LLMs with immediate, widespread applicability across all conversational AI deployments.
Paper 1 introduces a new, principled evaluation metric (Synergistic Faithfulness) addressing a fundamental flaw in current VLM explainability evaluation (cross-modal redundancy causing metric collapse). It is methodologically grounded (Shapley/Harsanyi interaction), demonstrates strong surrogate validity and large compute gains, and is evaluated broadly (multiple explainers, architectures, datasets). Its impact spans VLM auditing, XAI methodology, and safety-critical deployment practices—highly timely given rapid VLM adoption. Paper 2 is useful and relevant to LLM safety/steering, but is more incremental (repurposing persona vectors) with narrower generality and fewer systems tested.
CLORE addresses a broadly impactful problem—improving reasoning efficiency in LLMs—with a novel content-level optimization framework that complements existing length-based methods. It demonstrates compatibility with multiple training algorithms (GRPO, DAPO, etc.) across multiple benchmarks, suggesting wide applicability. Paper 1, while offering interesting mechanistic insights into sycophancy via persona vectors, is narrower in scope and incremental relative to existing steering methods. The reasoning efficiency problem tackled by Paper 2 is more timely given the rapid deployment of reasoning models, and its framework is likely to see broader adoption.
AutoRPA addresses a practical and broadly applicable problem—bridging LLM-based agents and traditional RPA for efficient GUI automation. It introduces a concrete framework with measurable improvements (82-96% token reduction) and has clear real-world applications in enterprise automation. Paper 1, while offering interesting mechanistic insights into sycophancy as a persona-level property, is narrower in scope, focused on interpretability of steering vectors with incremental findings. Paper 2's combination of novelty, practical utility, and cross-field relevance (NLP, software engineering, HCI) gives it higher potential impact.
Paper 2 likely has higher impact: it targets an urgent, widely relevant safety/alignment failure mode (sycophancy) in mainstream LLM deployments, proposes a practical mitigation using off-the-shelf persona vectors (no labeled sycophancy data), and reports strong performance with less accuracy tradeoff than CAA. Its finding that sycophancy is not a single steerable direction but relates to broader persona geometry could influence steering, evaluation, and interpretability work across many models and applications. Paper 1 is novel but narrower, focused on specialized spatial/recursive reasoning benchmarks and analysis tooling.