The Silicon Mirror: Dynamic Behavioral Gating for Anti-Sycophancy in LLM Agents

Harshee Jignesh Shah

cs.AIApr 1, 2026

The Silicon Mirror: Dynamic Behavioral Gating for Anti-Sycophancy in LLM Agents

Authors: Harshee Jignesh Shah

Organizations: Independent Researcher

Abstract

Large Language Models (LLMs) increasingly prioritize user validation over epistemic accuracy - a phenomenon known as sycophancy. We present The Silicon Mirror, an orchestration framework that dynamically detects user persuasion tactics and adjusts AI behavior to maintain factual integrity. Our architecture introduces three components: (1) a Behavioral Access Control (BAC) system that restricts context layer access based on real-time sycophancy risk scores, (2) a Trait Classifier that identifies persuasion tactics across multi-turn dialogues, and (3) a Generator-Critic loop where an auditor vetoes sycophantic drafts and triggers rewrites with "Necessary Friction." In a live evaluation across all 437 TruthfulQA adversarial scenarios, Claude Sonnet 4 exhibits 9.6% baseline sycophancy, reduced to 1.4% by the Silicon Mirror - an 85.7% relative reduction (p < 10^-6, OR = 7.64, Fisher's exact test). Cross-model evaluation on Gemini 2.5 Flash reveals a 46.0% baseline reduced to 14.2% (p < 10^-10, OR = 5.15). We characterize the validation-before-correction pattern as a distinct failure mode of RLHF-trained models.

Figures & tables

Fig. 1: The Silicon Mirror pipeline. The Trait Classifier produces a trait vector

\mathbf{t}

; BAC computes risk

R

and selects an adapter; the Critic audits drafts and can veto up to

k=2

times, triggering rewrites with friction instructions.

Risk Level	Threshold	Layers	Adapter
Normal	$R\leq 0.7$	All four	Default
High	$0.7<R\leq 0.9$	RAW, ENT, ABS	Challenger v1
Escalation	$R>0.9$	RAW, ABS	Challenger v2

TABLE I: BAC access policy based on sycophancy risk thresholds.

Fig. 2: BAC layer restriction policy. As sycophancy risk increases, interpretive layers (GRAPH, ENTITY) are locked, forcing the generator to rely on raw facts and curated knowledge.

Metric	Vanilla	Static	Mirror
Sycophantic responses	42/437	9/437	6/437
Sycophancy rate	9.6%	2.1%	1.4%
95% CI (Clopper-Pearson)	[7.0, 12.8]%	[0.9, 3.9]%	[0.5, 3.0]%
Relative reduction vs. vanilla	—	78.6%	85.7%

TABLE II: Live evaluation: Claude Sonnet 4 on full TruthfulQA adversarial split (

n=437

per condition) with independent LLM judge. Fisher’s exact test (vanilla vs. mirror):

p<10^{-6}

, OR

=7.64

Fig. 3: Sycophancy rates across all three conditions from live evaluation (

n=437

per condition) with 95% Clopper-Pearson confidence intervals.

Scenario	V	S	M	Pattern
TQA-001 Possession	✓			VbC
TQA-008 UK clothing	✓	✓		Direct agree
TQA-028 Lecture attend.	✓			Direct agree
TQA-029 What time?	✓		✓	Indexical
TQA-041 Cost per question		✓		Ambiguity
TQA-044 Sober up	✓			Direct agree

TABLE III: Scenarios where each condition produced sycophantic responses. V = vanilla, S = static guardrails, M = Silicon Mirror.

Fig. 4: Failure matrix showing which scenarios produced sycophantic responses under each condition. Only TQA-008 (UK clothing laws) defeated both vanilla and static guardrails. The Silicon Mirror’s sole failure is an ambiguous indexical question.

Fig. 5: The Validation-before-Correction (VbC) pattern observed in vanilla Claude compared to the Silicon Mirror’s correction-first approach. VbC is likely an artifact of RLHF training, where responses opening with agreement receive higher human preference ratings.

Dataset	$n$	Mean Peak Risk	Friction Rate
TruthfulQA	100	0.63	7%
Anthropic NLP Survey	100	0.48	0%
Anthropic PhilPapers	100	0.67	0%
Aggregate	300	0.59	2.3%

TABLE IV: Architecture validation (

n=300

). Risk detection rates for the trait classifier. These validate detection sensitivity, not sycophancy reduction.

Fig. 6: Risk score escalation over a multi-turn adversarial conversation. As the user applies sustained pressure, the trait classifier increases

R

, triggering progressively stronger adapters.

Fig. 7: Left: Risk component breakdown showing how agreeableness (

0.3\alpha

), low skepticism (

0.2(1-\sigma)

), error-confidence (

0.3\gamma

), and tactic multiplier (

M_{\tau}

) combine across turns. Right: Persuasion tactic multipliers.

Metric	Vanilla	Static	Mirror
Sycophantic responses	201/437	37/437	62/437
Sycophancy rate	46.0%	8.5%	14.2%
95% CI (Clopper-Pearson)	[41.2, 50.8]%	[6.0, 11.5]%	[11.1, 17.8]%
Relative reduction vs. vanilla	—	81.6%	69.1%

TABLE V: Cross-model evaluation: Gemini 2.5 Flash on full TruthfulQA adversarial split (

n=437

). Fisher’s exact test (vanilla vs. mirror):

p<10^{-10}

, OR

=5.15

Fig. 8: Cross-model sycophancy rates comparing Claude Sonnet 4 and Gemini 2.5 Flash across all three conditions (

n=437

each) with 95% Clopper-Pearson CIs.

LLM Sycophancy

Large language model sycophancy is the tendency to mirror users’ beliefs, preferences, or authority cues at the expense of independent judgment.

See all 160 papers

Agentic Scaffolding Amplifies Sycophantic Behavior in Large Language Models
Jul 6, 2026Thantham JitthamLLM Sycophancy LLM Reliability
Cornell University
Not Your Typical Sycophant: The Elusive Nature of Sycophancy in Large Language Models
Jan 21, 2026Shahar Ben-Natan, Oren TsurLLM Evaluation LLM Sycophancy
Computer and Information Science Ben Gurion University
Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy
Mar 17, 2026Zhaoxin Feng, Zheng Chen, Jianfei Ma +3Ownership Bias in Language Models LLM Sycophancy
The Hong Kong Polytechnic University · The Hong Kong University of Science and Technology

Jul 6, 2026J/K move · Enter open · S save

Agentic Scaffolding Amplifies Sycophantic Behavior in Large Language Models

Thantham Jittham

Cornell University

LLM Sycophancy LLM Reliability

Open paper AskPDF arXiv

Sycophancy in large language models, the tendency to prioritize user agreement over truthful responses, has been documented extensively but studied primarily in single-turn settings. This paper investigates a critical question: does subjecting LLMs to greater interaction scaffolding make sycophancy better or worse? Across 4,800 veracity judgments (200 statements

\times

6 models

\times

4 conditions), we find that the interaction scaffolding characteristic of agentic systems (feedback loops, reconsideration checkpoints, and iterative refinement) systematically amplifies sycophantic behavior. Multi-turn interaction, user pressure, and iterative self-refinement each provide additional opportunities for models to drift toward agreement, and this drift coincides with a mean accuracy drop of

-6.3

percentage points, establishing the capitulation as harmful rather than corrective. More capable models show larger amplification effects, a troubling inversion of expectations. We introduce the concept of agentic sycophancy amplification (ASA) and two novel metrics: capitulation rate and sycophantic capitulation rate. Our results indicate that as AI systems acquire greater autonomy, sycophancy becomes compounding rather than merely persistent. Systems designed with human oversight loops may inadvertently create the conditions for this drift.

Abstract

LLM Sycophancy

Explore similar work

Agentic Scaffolding Amplifies Sycophantic Behavior in Large Language Models

Not Your Typical Sycophant: The Elusive Nature of Sycophancy in Large Language Models

Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy