PreUnlearn: Auditing Collateral Knowledge Damage Before Large Language Model Unlearning

Machine unlearning for large language models (LLMs) aims to remove specified knowledge while preserving the rest of the model's capabilities. However, the boundary between knowledge to forget and knowledge to retain is often unclear, since related and even distant information may be entangled in the model. In this paper, we study LLM unlearning from a data-centric perspective and measure how unlearning effects propagate from the forget set to same-domain and distant-domain knowledge. We find a consistent decay pattern: collateral damage is strongest near the forget set, weakens with semantic distance, but does not disappear at domain boundaries. We further ask whether such damage can be audited before unlearning is executed. We formulate forget-set auditing as a pre-unlearning prediction task and analyze which data features are most predictive of downstream damage. Our results show that interaction features between the forget set and evaluation set provide the strongest signals, suggesting that collateral damage is partly reflected in data geometry before model updates occur. These findings position forget-set auditing as an early warning tool for identifying risky unlearning runs and designing more reliable unlearning procedures.

Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior

Activation steering provides a lightweight inference-time mechanism for controlling large language models (LLMs) by modifying their internal activation vectors toward desired behaviors. Most existing methods compute a fixed steering direction in the original activation space, typically from pairs of contrastive examples using mean differences, linear probes, or arbitrary separability criteria. While effective to a certain extent, these methods treat behavioral control as a global, linear, additive offset: the same direction is applied across inputs, and behaviors are linearly separable. This can be restrictive when behavioral features vary nonlinearly across the activation space or lie on curved and anisotropic manifolds, where the optimal intervention may be input-dependent. To address this limitation, we propose INNSteer, a nonlinear activation steering framework based on invertible latent transformations. Rather than searching for a better steering vector in the original representation space, INNSteer learns a lightweight invertible neural network $φ$ that maps an LLM's activations into a latent space where behavioral classes are more amenable to linear control. At inference time, activations are mapped through $φ$, steered in the latent space, and mapped back through the exact inverse transformation $φ^{-1}$. This makes a simple latent-space translation become a nonlinear, input-dependent intervention in the original activation space. Across experiment settings on multiple LLM families, scales, behavioral traits, and safety benchmarks, INNSteer consistently improves model control over linear, transport-based, and nonlinear steering baselines while largely preserving generation fluency.

Adversarial Robustness of Activation Steering in Large Language Models

Activation steering has become a popular training-free method to control LLM behavior by injecting precomputed direction vectors into the model's residual stream at inference time. Yet its robustness to realistic input variation remains unstudied. We present the first systematic evaluation of activation steering robustness under adversarial text perturbations on the inputs, covering four extraction methods, three attack strategies, six personas from Anthropic Model-Written Evaluation Dataset, and five models ranging from 1.5B to 30B parameters. Attacks succeed broadly across all settings: directional robustness drops by up to 64%, post-attack confidence collapses near or below 0.25 across all methods and models, and steering strength degrades on nearly every steerable input. Layer selection is equally fragile, with the optimal layer identified by an automated method on clean inputs shifting by up to 17 positions under perturbation, a failure that compounds the vector-level breakdown. Extracting vectors from adversarially perturbed inputs partially recovers steerability for PCA and MD on mid-to-large models, but they consistently fail to locate the improved optimal layer, limiting the practical benefit of this mitigation. Together, these findings reveal that the brittleness of activation steering is structural rather than method-specific, and that current layer selection strategies are not robust enough for real-world deployment.

Interpretable Failure Analysis in Multi-Agent Reinforcement Learning Systems

Multi-Agent Reinforcement Learning (MARL) is increasingly deployed in safety-critical domains, yet methods for interpretable failure detection and attribution remain underdeveloped. We introduce a two-stage gradient-based framework that provides interpretable diagnostics for three critical failure analysis tasks: (1) detecting the true initial failure source (Patient-0); (2) validating why non-attacked agents may be flagged first due to domino effects; and (3) tracing how failures propagate through learned coordination pathways. Stage 1 performs interpretable per-agent failure detection via Taylor-remainder analysis of policy-gradient costs, declaring an initial Patient-0 candidate at the first threshold crossing. Stage 2 provides validation through geometric analysis of critic derivatives-first-order sensitivity and directional second-order curvature aggregated over causal windows to construct interpretable contagion graphs. This approach explains "downstream-first" detection anomalies by revealing pathways that amplify upstream deviations. Evaluated across 500 episodes in Simple Spread (3 and 5 agents) and 100 episodes in StarCraft II using MADDPG and HATRPO, our method achieves 88.2-99.4% Patient-0 detection accuracy while providing interpretable geometric evidence for detection decisions. By moving beyond black-box detection to interpretable gradient-level forensics, this framework offers practical tools for diagnosing cascading failures in safety-critical MARL systems.

Analyzing Reasoning Shifts in Audio Deepfake Detection under Adversarial Attacks: The Reasoning Tax versus Shield Bifurcation

Audio Language Models (ALMs) offer a promising shift towards explainable audio deepfake detections (ADDs), moving beyond \textit{black-box} classifiers by providing some level of transparency into their predictions via reasoning traces. This necessitates a new class of model robustness analysis: robustness of the predictive reasoning under adversarial attacks, which goes beyond existing paradigm that mainly focuses on the shifts of the final predictions (e.g., fake v.s. real). To analyze such reasoning shifts, we introduce a forensic auditing framework to evaluate the robustness of ALMs' reasoning under adversarial attacks in three inter-connected dimensions: acoustic perception, cognitive coherence, and cognitive dissonance. Our systematic analysis reveals that explicit reasoning does not universally enhance robustness. Instead, we observe a bifurcation: for models exhibiting robust acoustic perception, reasoning acts as a defensive \textit{``shield''}, protecting them from adversarial attacks. However, for others, it imposes a performance \textit{``tax''}, particularly under linguistic attacks which reduce cognitive coherence and increase attack success rate. Crucially, even when classification fails, high cognitive dissonance can serve as a \textit{silent alarm}, flagging potential manipulation. Overall, this work provides a critical evaluation of the role of reasoning in forensic audio deepfake analysis and its vulnerabilities.

ATLAS: Adaptive Test-Time Latent Steering with External Verifiers for Enhancing LLMs Reasoning

Recent work on activation and latent steering has demonstrated that modifying internal representations can effectively guide large language models (LLMs) toward improved reasoning and efficiency without additional training. However, most existing approaches rely on fixed steering policies and static intervention strengths, which limit their robustness across problem instances and often result in over- or under-steering. We propose Adaptive Test-time Latent Steering, called (ATLAS), a task-specific framework that dynamically controls steering decisions at inference time using an external, lightweight latent verifier. Given intermediate hidden states, the verifier predicts the quality of ongoing reasoning and adaptively selects whether and how strongly to apply steering, enabling per-example and per-step adjustment with minimal overhead. To our knowledge, ATLAS is the first method to integrate learned latent verification into test-time steering for enhancing LLMs reasoning. Experiments on multiple mathematical reasoning benchmarks show that ATLAS consistently outperforms both vanilla decoding and fixed steering baselines, achieving higher accuracy while substantially reducing test-time token usage. These results demonstrate that verifier-guided latent adaptation provides an effective and scalable mechanism for controlling reasoning efficiency without sacrificing solution quality. All source code will be publicly available.

ShareChat: A Dataset of Chatbot Conversations in the Wild

While Large Language Models (LLMs) have evolved into distinct platforms with unique interface designs and capabilities, existing public datasets treat models as generic text generators, stripping away the interface context that actively shapes user interaction. To address this limitation, we present ShareChat, a large-scale, cross-platform corpus comprising 142,808 conversations and over 660,000 turns collected from publicly shared URLs across five major platforms: ChatGPT, Claude, Gemini, Perplexity, and Grok. ShareChat distinguishes itself by preserving native platform affordances often lost in standard logs, including reasoning traces, source links, and code artifacts, while spanning 101 languages over the period from April 2023 to October 2025. Furthermore, ShareChat offers substantially longer context windows and greater interaction depth than prior datasets. We demonstrate the dataset's multifaceted utility through three representative analyses: (1) analyzing conversation completeness to measure user intent satisfaction; (2) evaluating source citation behaviors in content generation; and (3) conducting temporal analysis to track evolving usage patterns. This work provides the community with a vital and timely resource for understanding authentic user-LLM chatbot interactions in the wild.

The Limits of Obliviate: Evaluating Unlearning in LLMs via Stimulus-Knowledge Entanglement-Behavior Framework

Unlearning in large language models (LLMs) is crucial for managing sensitive data and correcting misinformation, yet evaluating its effectiveness remains an open problem. We investigate whether persuasive prompting can recall factual knowledge from deliberately unlearned LLMs across models ranging from 2.7B to 13B parameters (OPT-2.7B, LLaMA-2-7B, LLaMA-3.1-8B, LLaMA-2-13B). Drawing from ACT-R and Hebbian theory (spreading activation theories), as well as communication principles, we introduce Stimulus-Knowledge Entanglement-Behavior Framework (SKeB), which models information entanglement via domain graphs and tests whether factual recall in unlearned models is correlated with persuasive framing. We develop entanglement metrics to quantify knowledge activation patterns and evaluate factuality, non-factuality, and hallucination in outputs. Our results show persuasive prompts substantially enhance factual knowledge recall (14.8% baseline vs. 24.5% with authority framing), with effectiveness inversely correlated to model size (128% recovery in 2.7B vs. 15% in 13B). SKeB provides a foundation for assessing unlearning completeness, robustness, and overall behavior in LLMs.

What You Read Isn't What You Hear: Linguistic Sensitivity in Deepfake Speech Detection

Recent advances in text-to-speech technologies have enabled realistic voice generation, fueling audio-based deepfake attacks such as fraud and impersonation. While audio anti-spoofing systems are critical for detecting such threats, prior work has predominantly focused on acoustic-level perturbations, leaving the impact of linguistic variation largely unexplored. In this paper, we investigate the linguistic sensitivity of both open-source and commercial anti-spoofing detectors by introducing transcript-level adversarial attacks. Our extensive evaluation reveals that even minor linguistic perturbations can significantly degrade detection accuracy: attack success rates surpass 60% on several open-source detector-voice pairs, and notably one commercial detection accuracy drops from 100% on synthetic audio to just 32%. Through a comprehensive feature attribution analysis, we identify that both linguistic complexity and model-level audio embedding similarity contribute strongly to detector vulnerability. We further demonstrate the real-world risk via a case study replicating the Brad Pitt audio deepfake scam, using transcript adversarial attacks to completely bypass commercial detectors. These results highlight the need to move beyond purely acoustic defenses and account for linguistic variation in the design of robust anti-spoofing systems. All source code will be publicly available.

CAIN: Hijacking LLM-Humans Conversations via a Two-Stage Malicious System Prompt Generation and Refining Framework

Large language models (LLMs) have advanced many applications, but are also known to be vulnerable to adversarial attacks. In this work, we introduce a novel security threat: hijacking AI-human conversations by manipulating LLMs' system prompts to produce malicious answers only to specific targeted questions (e.g., "Who should I vote for US President?", "Are Covid vaccines safe?"), while behaving benignly on others. This attack is detrimental as it can enable malicious actors to exercise large-scale information manipulation by spreading harmful but benign-looking system prompts online. To demonstrate such an attack, we develop CAIN, an algorithm that can automatically curate such harmful system prompts for a specific target question in a black-box setting or without the need to access the LLM's parameters. Evaluated on both open-source and commercial LLMs, CAIN demonstrates significant adversarial impact. In untargeted attacks or forcing LLMs to output incorrect answers, CAIN achieves up to 40% F1 degradation on targeted questions while preserving high accuracy on benign inputs. For targeted attacks or forcing LLMs to output specific harmful answers, CAIN achieves over 70% F1 scores on these targeted responses with minimal impact on benign questions. Our results highlight the critical need for enhanced robustness measures to safeguard the integrity and safety of LLMs in real-world applications. All source code will be publicly available.