When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception

• URL: http://arxiv.org/abs/2605.30381v1
• Date: Thu, 28 May 2026 01:20:06 GMT
• Title: When LLMs Learn to Be Consistently Wrong: A Multi-Model Study of Linear Representations of Synthetic Deception
• Abstract summary: Deceptive alignment, in which models maintain accurate internal representations while deliberately producing false outputs, remains a central challenge in AI safety.<n>We introduce a multi-model paradigm in which honest and deceptive variants of five transformer models are fine-tuned using LoRA on the same question distribution.<n> Linear probes trained on mean-pooled hidden states detect synthetic dishonesty with near-perfect AUC.<n>Mechanistic analysis of Fisher Discriminant Ratio, effective rank, centroid geometry, directional calibration, and calibration (ECE) reveals two stability: representational collapse in Pythia/Llama/Qwen versus high-dimensional preservation in
• Score: 0.0
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Deceptive alignment, in which models maintain accurate internal representations while deliberately producing false outputs, remains a central challenge in AI safety. While strategic deception is the primary long-term concern, synthetic dishonesty - induced via direct optimization on incorrect answers - provides a controlled testbed for studying the representational basis of learned deception. We introduce a multi-model paradigm in which honest and deceptive variants of five transformer models (Pythia-1.4B, Gemma-2-2B/9B, Qwen2.5-7B, Llama-3.1-8B) are fine-tuned using LoRA on the same question distribution. Linear probes trained on mean-pooled hidden states detect synthetic dishonesty with near-perfect AUC (greater than or equal to 0.99) as early as layers 1-3 in four architectures, while Pythia-1.4B reaches a peak of 0.705. Logistic regression probes consistently match or outperform MLP probes, supporting the Linear Representation Hypothesis. Probes trained on TruthfulQA generalize with near-zero loss (Delta AUC approx. 0) to held-out MMLU subjects. Late-layer representations show strong robustness to Gaussian noise, with Gemma-2 models exhibiting exceptional stability. Mechanistic analysis of Fisher Discriminant Ratio, effective rank, centroid geometry, directional stability, cross-domain alignment, and calibration (ECE) reveals two regimes: representational collapse in Pythia/Llama/Qwen versus high-dimensional preservation in Gemma-2. Across all models, the dishonesty direction consolidates progressively in deeper layers, with optimal calibration (ECE less than 0.01 except Pythia) achievable in layers 1-4. These results demonstrate that robust, domain-invariant dishonesty representations can be rapidly entrenched via modest supervised fine-tuning, with implications for activation-based monitoring.

Related papers

• IFW-BLS: Dual-Robust Broad Learning System with Intuitionistic Fuzzy Wave Loss [7.2201008524483585]
  Broad Learning System is an efficient randomized learning model that expands network width through feature and enhancement nodes.<n>Its standard least-squares training, however, is vulnerable in two different ways: (i) large residuals caused by noise, outliers, or corrupted labels can dominate the objective, and (ii) all samples are treated as equally reliable even when some lie in ambiguous or locally conflicting regions.<n>This paper proposes IFW-BLS, an Intuitionistic Fuzzy Wave Broad Learning System that addresses these two sources of one optimization model.
  arXiv  Detail & Related papers  (2026-09-02T10:42:24Z)
• Perception, Verdict, and Evolution: Hindsight-Driven Self-Refining Forensics Agent for AI-Generated Image Detection [80.15317858033534]
  ForeAgent is an agentic forensics framework for AI-generated image detection with iterative self-evolution.<n>ForeAgent achieves state-of-the-art performance on the Chameleon benchmark, reaching 82.18% accuracy.<n>ForeAgent produces more consistent and causally grounded reasoning compared to GPT-5 and GPT-5-mini.
  arXiv  Detail & Related papers  (2026-06-25T02:59:33Z)
• Seirênes: Adversarial Self-Play with Evolving Distractions for LLM Reasoning [56.48520300004217]
  We present Seirnes, a self-play RL framework that transforms contextual interference into an internal training signal.<n>A single model is trained to both construct plausible yet distracting contexts that expose its own reasoning blind spots.<n>By pitting these competing objectives against each other, Seirnes compels the model to move beyond superficial pattern matching.
  arXiv  Detail & Related papers  (2026-05-12T06:58:35Z)
• Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMs [5.478971182058342]
  Safety evaluations almost always study prompt-based personas alone.<n>We show this is incomplete: prompting and activation steering expose *different*, architecture-dependent vulnerability profiles.
  arXiv  Detail & Related papers  (2026-04-13T07:34:02Z)
• DOC-GS: Dual-Domain Observation and Calibration for Reliable Sparse-View Gaussian Splatting [80.43237927269575]
  In this paper, we revisit sparse-view 3DGS reconstruction from a new perspective.<n>We identify the core challenge as the unobservability of Gaussian primitive reliability.<n>Motivated by this observation, we propose a unified Dual-domain Observation and geometrics in rendered images framework.
  arXiv  Detail & Related papers  (2026-04-08T07:01:24Z)
• The Geometry of Robustness: Optimizing Loss Landscape Curvature and Feature Manifold Alignment for Robust Finetuning of Vision-Language Models [33.10473765967503]
  Generalization-preserving methods retain ID/OOD performance but leave models vulnerable to adversarial attacks.<n>Our key insight is that the robustness trade-off stems from two geometric failures: sharp, anisotropic minima in parameter space and unstable feature representations that deform under perturbation.<n>We propose GRACE, a unified fine-tuning framework that jointly regularizes the parameter-space curvature and feature-space invariance for VLMs.
  arXiv  Detail & Related papers  (2026-03-28T05:22:00Z)
• Information Fidelity in Tool-Using LLM Agents: A Martingale Analysis of the Model Context Protocol [69.11739400975445]
  We introduce the first theoretical framework for analyzing error accumulation in Model Context Protocol (MCP) agents.<n>We show that cumulative distortion exhibits linear growth and high-probability deviations bounded by $O(sqrtT)$.<n>Key findings include: semantic weighting reduces distortion by 80%, and periodic re-grounding approximately every 9 steps suffices for error control.
  arXiv  Detail & Related papers  (2026-02-10T21:08:53Z)
• MSign: An Optimizer Preventing Training Instability in Large Language Models via Stable Rank Restoration [48.446476072756276]
  Training instability remains a critical challenge in large language model pretraining.<n>We study training failures in a 5M NanoGPT model scaled via $$P.<n>We propose MSign, a new norm that periodically applies matrix sign operations to restore stable rank.
  arXiv  Detail & Related papers  (2026-02-02T07:18:45Z)
• Distributional Clarity: The Hidden Driver of RL-Friendliness in Large Language Models [50.99097734404912]
  We show that RL-friendly models exhibit intra-class compactness and inter-class separation in their probability assignments to correct vs. incorrect responses.<n>Experiments across six mathematical benchmarks show consistent improvements across all model families, with gains up to 5.9 points on AIME24.
  arXiv  Detail & Related papers  (2026-01-11T13:34:44Z)
• Mitigating LLM Hallucination via Behaviorally Calibrated Reinforcement Learning [32.32593439144886]
  Behavior-calibrated reinforcement learning allows smaller models to surpass frontier models in uncertainty quantification.<n>Our model's log-scale Accuracy-to-Hallucination Ratio gain (0.806) exceeds GPT-5's (0.207) in a challenging in-domain evaluation.
  arXiv  Detail & Related papers  (2025-12-22T22:51:48Z)
• RefineBench: Evaluating Refinement Capability of Language Models via Checklists [71.02281792867531]
  We evaluate two refinement modes: guided refinement and self-refinement.<n>In guided refinement, both proprietary LMs and large open-weight LMs can leverage targeted feedback to refine responses to near-perfect levels within five turns.<n>These findings suggest that frontier LMs require breakthroughs to self-refine their incorrect responses.
  arXiv  Detail & Related papers  (2025-11-27T07:20:52Z)
• Parrot: Persuasion and Agreement Robustness Rating of Output Truth -- A Sycophancy Robustness Benchmark for LLMs [0.0]
  PARROT (Persuasion and Agreement Robustness Rating of Output Truth) is a robustness focused framework designed to measure the degradation in accuracy under social pressure exerted on users.<n>We evaluate 22 models using 1,302 MMLU-style multiple-choice questions across 13 domains and domain-specific authority templates.
  arXiv  Detail & Related papers  (2025-11-21T13:01:28Z)
• SEAL: Steerable Reasoning Calibration of Large Language Models for Free [58.931194824519935]
  Large Language Models (LLMs) have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism.<n>Recent studies reveal substantial redundancy in the CoT reasoning traces, which negatively impacts model performance.<n>We introduce SEAL, a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains.
  arXiv  Detail & Related papers  (2025-04-07T02:42:07Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.