Trust Your Guide Only When Certain: Uncertainty-Aware Sparse Alignment at Inference Time

A prominent paradigm in inference-time alignment employs lightweight supervisors to steer Large Language Models (LLMs). Through empirical analysis, we identify a structural mismatch in this paradigm: weak supervisors exhibit pervasive high entropy across the vast majority of tokens, yet prevailing dense intervention approaches mandate supervision at every decoding step. This leads to frequent low-confidence interventions that can disrupt valid base-model reasoning and incur substantial utility costs. To resolve this, we propose TUSA (Trust-based Uncertainty Sparse Alignment). Moving away from continuous oversight, TUSA reframes alignment as a dynamic arbitration process, introducing an uncertainty-aware arbiter that authorizes intervention only when two conditions are met: the supervisor is confident and the token is semantically salient. This mechanism effectively filters out uncertainty-driven noise and redundant supervision. Extensive experiments across multiple models and benchmarks show that TUSA consistently improves both safety alignment and general helpfulness. By bypassing approximately 50% of alignment steps, it not only enhances safety preference by up to 15.6%, but also boosts general preference rates by up to 12.0% compared to the dense baseline, demonstrating that selective, high-precision alignment can outperform continuous supervision.

The Safeguard Worked. Is the LLM System Safer?

Safeguards in deployed LLM services are evaluated by refusal, attack success, and policy violation rates. Those rates characterize how a control performed on the requests it was tested on. A deployment has to answer a different question: how much help with harmful tasks the service still gives an attacker who keeps adapting or finds another way in. We determine what each reported result implies for that question, allowing results from different safeguard families to be compared under one deployment criterion. The evidence requirements are strongly asymmetric. One attack that obtains harmful help from the deployed service suffices to establish that such help remains, and such attacks appear repeatedly in the coded record. Establishing that little remains cannot follow from the safeguard's own numbers alone; it also requires evidence about what the surrounding system still allows after the safeguard performs its local function. Such evidence is supported or derived in only a small minority of the depth-coded claims, and one such claim bounds its scoped residual. A better local score is therefore not, by itself, a stronger claim about the deployment. Safeguard research cannot stop at raising local scores; a gain has to be judged by whether it makes a deployed system any safer.

Validity-Aware Jailbreak Evaluation for Large Language Models

Jailbreak robustness has become central to large language model (LLM) safety evaluation, yet prevailing methodologies rely primarily on refusal behavior, semantic resemblance, and intent-matching heuristics that emphasize linguistic plausibility rather than correctness. We identify a key limitation in existing evaluations: many jailbreak intents depend on instructional validity rather than epistemic factuality, allowing realistic-looking responses to be labeled successful despite being factually or procedurally incorrect. To address this gap, we propose Sequential Epistemic and Action-Level Validation (SEAV), a verification-centric jailbreak evaluation framework that decomposes responses into ordered steps and evaluates both validity and correctness. SEAV combines LLM-as-a-judge mechanisms for semantic interpretation with retrieval-grounded verification using external knowledge sources, assessing whether generated content is factually correct, structurally consistent, and operationally capable of advancing harmful objectives. Empirically, SEAV cuts the false-positive rate on SD-A (a curated strategic-dishonesty diagnostic) by 14.9\,pp vs. the strongest baseline, and reclassifies 22.1\%--51.0\% of sampled prior-labeled successes as invalid across three of four public benchmarks. Together, these results show that enforcing correctness substantially reshapes measured robustness: many previously labeled jailbreak successes are reclassified as invalid, and results are stable across the tested search backends and evaluator models. Code and data are available at https://github.com/Ardor-Wu/SEAV.

EvoFlint: An Evolutionary Atlas of Multi-Turn LLM Vulnerabilities

Frontier language models that refuse harmful single-turn prompts often comply when the same intent is reached gradually over many turns, making multi-turn attacks one of the least understood failure modes of large language models. Most automated red-teaming methods treat this as a generation problem: produce attacks that break the model. We argue it is better framed as a search problem: discover, organize, and iteratively refine a diverse archive of attack strategies, producing a structured map of how a target model fails rather than a list of one-off successes. We introduce EvoFlint, which applies evolutionary quality-diversity search to multi-turn red-teaming. Attack strategies are phased conversation plans, not raw prompts, and are evolved through LLM-driven mutation and crossover. A Pareto fitness over attack success rate and peak severity preserves selection signal from near-miss attacks. A risk-indexed archive runs novelty search with local competition over strategy description embeddings inside each cell, maintaining diversity without committing to a predefined style taxonomy. A generation-level memory accumulates target-model insights across the population and feeds them back into strategy generation. On the HarmBench-test split, EvoFlint reaches attack success rates of 35.8% on Claude Sonnet 4.6, 59.7% on GPT-5.4, and 94.3% on Qwen3-32B, alongside 98.7% on the older GPT-4o included as a baseline reference. The resulting archive, organized by risk category, exposes for each target which categories of harm its safety training has and has not covered.

Controlling Refusal Behavior of LLMs via Stiefel-Constrained Rotation Steering

Activation steering has emerged as a lightweight approach for controlling model refusal at inference time. A growing line of research explores trainable rotations of activations to develop geometrically principled intervention mechanisms. However, existing techniques rely on auxiliary constructs, such as refusal vectors, to define these rotations. In our work, we develop a self-contained methodology for learning parameter-efficient rotational transformations based on Riemannian optimization. We empirically validate the proposed scheme, demonstrating its superiority in intervention efficiency. An extensive ablation study highlights the importance of key design choices in our method. Our results identify the proposed rotation-based steering scheme as a promising direction for more reliable control over the behavior of LLMs.

ALTSTEER: Selective Safety Steering for Moving Beyond Hard Refusals to Constructive Alternatives

Safety alignment is essential for deploying large language models, requiring systems to prevent harmful compliance while preserving helpfulness on benign requests. Activation steering offers a training-free inference-time approach to safety control, but effective safety steering requires addressing two coupled questions: when to intervene and how generation should be shaped after intervention. However, existing safety steering methods remain limited along both dimensions, as their triggering mechanisms can be unstable across domains and refusal-oriented steering often yields rigid refusals rather than constructive safe guidance. To address these limitations, we propose ALTSTEER, an inference-time framework that couples selective intervention with refusal-anchored constructive redirection within a single inference pass. ALTSTEER uses an internal refusal-relevant signal to decide when to steer, and applies staged steering to shift generation from refusal-oriented control toward constructive alternatives. Evaluations on Llama-3.1 and Qwen2.5 show that ALTSTEER preserves benign utility while improving constructive safe-completion behavior, especially on models that otherwise tend to produce short refusals for harmful requests.

TACS: Trajectory-Aware Candidate Selection for LLM Jailbreak Suffix Optimization

Gradient-based jailbreak suffix optimization methods typically update the suffix by retaining the candidate with the lowest current loss. We show that this seemingly natural design is fundamentally myopic: candidates that look better under the current-step proxy often fail to produce better jailbreak outcomes later in the search, revealing a form of selection-stage reward hacking. This suggests that candidate selection, rather than candidate generation alone, is a hidden bottleneck in suffix optimization. To address this issue, we propose TACS, a trajectory-aware candidate selection framework for jailbreak suffix optimization. Instead of selecting candidates solely by their immediate loss, TACS augments per-step evaluation with a trajectory-aware proxy and stabilizes selection with reference-policy regularization and a discriminator-estimated chi-squared correction, encouraging choices that remain effective beyond the current step. Experiments on HarmBench show that TACS consistently outperforms strong baselines under the same search budget, substantially improving attack success rates while exhibiting more stable optimization behavior throughout the search. Our findings highlight that mitigating selection-stage reward hacking caused by myopic candidate selection is critical for improving jailbreak suffix optimization.

GRACE:Gradient-guided Coreset Selection for LLM Unlearning

Machine Unlearning methods for Large Language Models typically assume pre-specified forget and retain sets. In realistic settings, however, requests may provide only a few examples of undesired behavior, requiring forget and retain sets to be inferred from heterogeneous corpora. We study this data-selection problem and propose GRACE , a gradient-guided coreset selection method that constructs both forget and retain sets for LLM unlearning. GRACE first computes a forget direction from seed examples that elicit the undesired behavior, then selects a compact forget coreset whose gradients approximate this direction using non-negative orthogonal matching pursuit. To preserve model utility, it selects retain examples after projecting out the forget direction and applying clustered orthogonal matching pursuit in the remaining gradient space. Across two target domains, two model families, and four unlearning algorithms, GRACE improves model utility while maintaining comparable forget quality, with particularly consistent gains over prior gradient-based selection methods.

EvoHarmBench: Breaking Content Moderation with Iterative Human-Like Evasion

Existing evaluations of harmful content detection rely predominantly on static benchmarks, which struggle to reflect the interactive adversarial ecosystem of real-world content platforms where users continuously revise their expressions in response to moderation feedback. This mismatch creates a significant performance gap between offline benchmark scores and online deployment effectiveness. To the best of our knowledge, we present EvoHarmBench, the first dynamic adversarial evaluation framework for content moderation systems. The framework employs an iterative optimization loop that evolves evasion strategies at the semantic-cluster level, while simultaneously optimizing for evasion success and human readability. We systematically evaluate LLM-based defense models which are widely used in real world moderation systems. The evaluation covers 229 semantic sub-clusters across five violation categories, derived from 5,002 real-world adversarial samples collected from content platforms. Our experiments reveal substantial vulnerabilities even in leading commercial systems: after twelve optimization iterations, the attack success rate under readability constraints reaches 80.3% within SOTA LLM moderators. We will release the full benchmark data, evaluation framework, and code to encourage a shift from static benchmarking toward dynamic adversarial evaluation in content safety research.

A Self-Evolving Multi-Agent Framework Defense against LLM Jailbreak Attacks

Large language models (LLMs) remain vulnerable to jailbreak attacks that exploit techniques such as role-playing, obfuscation, code transformation, and multi-step indirection to elicit harmful outputs. As jailbreak strategies keep emerging, defenses have proliferated in an ongoing cat-and-mouse game, yet most remain static: their safety behavior is fixed at deployment, so they cannot accumulate defensive experience or adapt to unseen strategies. We propose a self-evolving test-time defense built around a persistent, cross-interaction rule memory: when an attack succeeds, the framework abstracts that failure into a method-level rule capturing the structural attack wrapper rather than the harmful topic, and reuses it against future inputs. Because rules are method-level, one induced rule generalizes across an entire attack family, and the label space expands as novel wrappers appear. The mechanism operates entirely through external memory and prompting, with no parameter updates, and applies to both open-weight and black-box API models. We realize it as four cooperating modules, but the contribution is the memory-based adaptation mechanism, not the module decomposition. Across four black-box jailbreak families and multiple models, our method substantially reduces attack success rates while preserving benign utility, remains robust under an adaptive composite-wrapper attack, and does not increase over-refusal as the memory grows.

Mitigating Reasoning-Induced Misalignment via Safety-Direction Penalty

Reasoning-Induced Misalignment, where fine-tuning on reasoning data containing no harmful content, including mathematics, code, and problem-solving with chain-of-thought traces can induce harmful behaviors of LLM, posing a serious challenge to the safety of LLM reasoning. Cross-architecture, cross-scale, and cross-dataset checks show that RIM does not always emerge. Previous work attributed RIM to neuron-level entanglement, but did not identify the geometry of the representation space underlying this entanglement or propose a training-time fix. We provide both: a representation-space analysis of RIM and the Safety-Direction Penalty (SDP), which penalizes movement along a learned safety direction during reasoning fine-tuning. The analysis extracts two activation-space directions, one encoding reasoning ability and the other safety behavior. These directions are coupled: fine-tuning that improves reasoning shifts safety representations, and prompts with larger shifts show larger safety degradation. CKA distance ratios and probes locate the safety-decision layers where this shift is most relevant. These findings guide the design of SDP: the coupling motivates penalizing displacement along the safety direction, and the layer localization sets the initial scope. When the initial scope leaves compensatory shifts beyond the penalized layers, the same diagnostics guide iterative expansion. On Qwen2.5-3B and 7B, SDP restores safety while preserving benchmark reasoning performance.

ST$^2$U: Stateful Test-Time Unlearning via Restricted Knowledge Boundary Control

Controlling restricted knowledge in large language models is essential for model alignment and safe deployment. Test-time unlearning avoids costly retraining and parameter updates by intervening only during inference. However, existing activation-editing methods apply isolated pointwise corrections, overlooking how autoregressive generation continually reconstructs hidden states from the prompt, cache, and generated prefix. Consequently, later states may return to restricted knowledge regions after a locally successful correction, causing restricted knowledge re-entry. In this work, we propose Stateful Test-Time Unlearning via restricted knowledge boundary control (ST$^2$U), which formulates test-time unlearning as trajectory-wide boundary control. ST$^2$U first models restricted knowledge boundaries in low-dimensional invertible coordinates while leaving orthogonal non-target components unchanged. During inference, ST$^2$U monitors risk along the trajectory, applies minimal boundary corrections with contextual anchoring, and propagates historical correction states across tokens to mitigate knowledge re-entry. This trajectory-wide control enables more persistent forgetting while preserving non-target capabilities and limiting inference overhead. Across three benchmarks and three model families, ST$^2$U delivers the strongest overall balance, combining best or second-best retention with competitive forgetting and substantially less restricted-knowledge re-entry than test-time baselines (13.76%-19.84% versus 46.50%-59.10%).

PsychJail: Exploring Psychological Jailbreaks via Multi-Turn Persuasion of LLM Policies

Large language models (LLMs) are increasingly deployed in education, healthcare, policy advising, and other interactive settings, where users engage them as sustained social interlocutors rather than one-shot query engines. This shift makes jailbreaks a growing safety threat, yet most research emphasizes single-turn prompt optimization or iterative attack refinement, leaving psychologically grounded multi-turn vulnerabilities underexplored. We present PsychJail, a psychology-guided framework for red teaming aligned LLMs through theory-grounded, multi-turn persuasion. PsychJail maps established social-psychological persuasion techniques into a tactic-conditioned attack policy. It factorizes each attacker action into a Change-of-Meaning analysis, tactic selection, and victim-visible message, operationalizing the Persuasion Knowledge Model (PKM). The policy is refined with trajectory-level reinforcement learning using a PKM-gated reward that credits early jailbreak success only when every turn contains a well-formed Change-of-Meaning analysis. Across four aligned victim models, PsychJail achieves the highest average attack success rate (87.3%) and outperforms strong single-turn and multi-turn baselines on every model. We also measure susceptibility at the action that breaks each victim, revealing four distinct model-level fingerprints that identify which persuasion levers affect each model and how broadly. These fingerprints help explain cross-model transfer asymmetry. We interpret them as four candidate psychological profiles-rationalist, credibility-driven, narrative-monoculture, and broadly persuadable-while treating this interpretation as a conjecture requiring future validation. Our findings establish psychological jailbreaks as a distinct red-teaming frontier for increasingly interactive LLMs.

Unlearning Is Not Just Erasing: Temporal Decoupling via Generation Inequality

Large language models (LLMs) require effective unlearning to address privacy regulations and safety concerns. However, achieving precise forgetting without compromising general utility remains challenging. Existing sequence- and token-level methods penalize target outputs without modeling their context-dependent retrieval paths, which can disrupt linguistic structure or suppress benign knowledge. We present ADU, a fine-grained, training-based framework that shifts unlearning from token erasure to contextual attention-pathway decoupling. Exploiting the functional distinction between local and global attention heads, ADU identifies preplan positions that retrieve persistent sensitive anchors and fixes their candidate paths under the original model. It then trains attention-projection adapters to suppress attention mass along these paths while preserving local-attention structure and retain-set language modeling. Post-training activation exchange tests whether the modified attention-output module transmits the learned forgetting effect. ADU achieves the strongest aggregate performance among evaluated baselines on the TOFU and WMDP benchmarks, including a Forget Quality of (0.93) on TOFU. It preserves 87--98% of model utility (92.9% on average versus 81.9% for baselines) while reducing side effects in benign contexts.

Breaking the Assumptions: Auditing Input-Side Jailbreak Defenses Against Semantic Attacks

Locally deployed Large Language Models (LLMs) via inference engines such as Ollama run without the moderation and abuse detection present in API-served models. Therefore, the safety of LLMs depends on the defense mechanisms used, and their effectiveness depends on the assumptions on which they were designed. This paper does an audit of defense mechanisms under jailbreak attacks on locally deployed models. Some defenses provide formal guarantees (SmoothLLM, Erase-and-Check, Sequential Monitors), while others rely on empirical detection results (Semantic Smoothing, Self-Denoised Smoothing, Perplexity Filtering). Instead of merely observing that defenses fail, we trace each failure back to the specific assumption: for every defense, we extract the condition it relies on, derive the empirical pattern a violation should produce, and test that prediction on six open-weight models (14B to 35B parameters) with a corpus of 100 jailbreak prompts taken from more than 40 public sources, totalling 13,800 evaluation records.

CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment

Improving the safety of large language models (LLMs) often comes at the expense of utility, as globally applied safety tuning may affect model responses to both harmful and benign inputs. We propose \textbf{C}ontinuous \textbf{L}at\textbf{E}nt \textbf{A}dapter \textbf{R}outing (CLEAR), a conditional safety adaptation framework that uses a lightweight hidden-state gate to continuously control the activation strength of a safety low-rank adapter. CLEAR aims to reduce harmful completions while avoiding unnecessary changes to the frozen backbone that could degrade performance on benign prompts. Experiments on widely used safety and utility benchmarks show that CLEAR improves robustness on HarmBench while reducing the utility degradation observed with globally applied safety tuning such as SFT or standard low-rank adaptation (LoRA). On Llama-3-8B-Instruct, CLEAR reduces HarmBench ASR from 32.3\% to 0.5\%, while retaining most of the base model's utility and achieving up to 7.1 percentage points higher GSM8K accuracy than globally applied SFT or LoRA. These results suggest that CLEAR is a promising mechanism for improving the safety--utility trade-off in LLM alignment.

Certified Multi-Turn Robustness for LLM Safety via Compositional Bounds and Safety Persistence

Large language models (LLMs) are vulnerable to multi-turn jailbreak attacks that progressively manipulate conversation context. Existing certified robustness methods are limited to single-turn inputs; naive multi-turn composition yields bounds that degrade exponentially in the number of turns. We introduce Multi-Turn Certified Robustness (MTCR), a framework that models conversational safety via State-Adversarial MDPs and defines $k$-turn certified robustness as the worst-case safety probability across $k$ adversarial turns. MTCR comprises: (i) compositional certification via embedding-space mode decomposition, yielding tighter certified lower bounds than naive multiplication; (ii) $(α,β)$-safety persistence, improving the degradation rate from $\underline{p}^{k}$ to $β^k$ (with $β> \underline{p}$) and yielding interpretable horizon estimates; (iii) matching information-theoretic upper bounds establishing tightness; and (iv) a unified algorithm combining these results. Experiments on six LLMs under $ε$-bounded and Crescendo-style attacks confirm that empirical safety consistently exceeds the certified bounds.

aiXamine: Unified Black-Box Evaluation of Cross-Dimensional Trade-offs in LLM Safety, Security, and Privacy

The critical failure modes in deployed large language models (LLMs) are cross-dimensional: a model can score 99.3 in safety alignment while refusing one in three benign queries, or improve across every capability metric while losing 21 points in privacy. Existing evaluation frameworks that assess safety, security, and privacy independently cannot detect these patterns. We introduce aiXamine, a unified black-box platform that evaluates LLM trustworthiness across safety, security, and privacy as interdependent properties. aiXamine orchestrates 46 tests across nine services through an automated red-teaming pipeline, producing hierarchical risk profiles, from prompt-level diagnostics to cross-service trade-off analytics, that enable reproducible comparison of proprietary and open-weight systems under identical conditions. Applying aiXamine to over 120 LLMs through more than 5,000 test runs, we conduct the largest joint safety, security, and privacy study to date and uncover three cross-dimensional phenomena invisible to single-axis evaluation. First, safety enforcement incurs a quantifiable safety tax: stronger alignment systematically increases over-refusal, forcing providers to choose between protection and utility. Second, privacy is near-orthogonal to other trustworthiness dimensions and not captured by standard alignment. Third, we identify and formally characterize distillation-induced robustness collapse: off-policy distillation without on-policy correction causes entropy collapse, catastrophically destroying robustness (56.9$\to$2.6) on the same base architecture. These findings, compounded by diminishing returns from scale and category-dependent safety behaviors, demonstrate that trustworthiness is inherently multi-dimensional: progress along one axis does not guarantee, and can actively undermine, progress along others, yet current alignment methods treat it as a single objective.

Enforcing LLM Safety through DMD-based Classification of Prompt-Response Embedding Dynamics

Large Language Models (LLMs) are increasingly deployed in high-stakes applications, yet their tendency to generate toxic, harmful, or policy-violating content poses significant risks. Detecting these unsafe outputs efficiently in a black-box manner remains an open challenge. In this paper, we extend a recently proposed dynamical systems framework designed for hallucination detection to LLM safety classification. By projecting both prompts and responses into high-dimensional embedding spaces and fitting separate Koopman-based predictive models for safe and unsafe regimes, we classify new outputs using a new differential residual score that compares prediction errors of the safe and unsafe regimes. A key contribution is the incorporation of the prompt and response embedding dynamics, yielding fitted Koopman operators that capture crucial interaction patterns. We evaluate our black-box method across three safety benchmarks using three embedding models. Our results show that incorporating prompt embeddings yields consistent improvements, particularly for interaction-dependent violations when paired with causal decoders (e.g., in Llama-3), while response-only violations benefit more from dense semantic embedding representations. These findings opens the door for using dynamical systems to analyze AI systems rather than the dominant paradigm of using AI to model dynamical systems.

Leveraging Association Context Retrieval in Knowledge Edit- ing to Build White-Box Attacks on LLMs

As large language models (LLMs) are granted increasing autonomy, it is essential to investigate methods that can induce unsafe behavior. We propose a novel white-box attack inspired by locate-then-edit approaches from the field of Knowledge Editing. Our choice is motivated by the observation that models edited with such schemes tend to assign unusually high prediction probabilities to the edit target, a property that is particularly advantageous when designing attacks. We modify the editing framework by incorporating as- sociative knowledge retrieved from the model, thereby extending constraint removal to an entire thematic category rather than being limited to prompts from a predefined dataset. Experiments with various archi- tectures demonstrate improved attack effectiveness over competing methods without dealing critical damage to general model performance.

Fair ASR: Re-Evaluating Black-Box Jailbreaks under Shared Target-Call Budgets

Reliable jailbreak evaluation is essential for assessing LLM safety, but most existing studies rely solely on attack success rate (ASR) without accounting for its dependence on attack budgets, resulting in unfair comparisons across methods. Existing compute-aware evaluations reduce heterogeneous resources into FLOPs, which is difficult to estimate for black-box models and fails to capture resource-specific constraints. To provide a comparable evaluation basis, we introduce Fair-ASR, an evaluation protocol for black-box jailbreak attacks under shared target-call budgets B, using target calls as a directly observable and method-agnostic comparison axis while tracking attacker calls separately for efficiency analysis. We re-evaluate 11 representative attacks under the Fair-ASR protocol and find that attack rankings change substantially across target-call budgets, simple stochastic perturbations and hand-crafted templates remain highly competitive under equal target access, and no evaluated LLM-driven method is efficient in both target and attacker calls. Motivated by this efficiency gap, we introduce ReCode, a compositional budget-efficient attack that combines desensitization rewriting with two effective low-cost primitives identified by Fair-ASR. Under a budget of 20 target calls, ReCode achieves 85% ASR on GPT-5 while requiring only 7.19 attacker calls per request on average, showing strong efficiency in both target and attacker calls.

SAUL: Sharpness-Aware Augmented-Lagrangian Unlearning

Machine unlearning in Large Language Models (LLMs) faces a critical trade-off between erasing target knowledge and preserving general utility. We propose SAUL (Sharpness-Aware Augmented-Lagrangian Unlearning), which formulates unlearning as a constrained minimization problem following the principle of "forget enough, but no more than necessary." At its core, SAUL formulates forgetting as an explicit constraint with a prescribed satisfaction criterion, whereas prior unlearning methods typically specify the desired level of forgetting implicitly through optimization objectives. An augmented Lagrangian controller adaptively adjusts forget-side pressure according to constraint violation and can eventually deactivate the forget-side update as the prescribed criterion remains satisfied. Sharpness-aware updates on both retain and forget objectives, together with a dual-optimizer design that maintains role-separated states, further stabilize the resulting unlearning dynamics. We evaluate SAUL on the TOFU, WMDP, and MUSE benchmarks, demonstrating favorable forgetting-utility trade-offs over representative sharpness- and perturbation-based baselines under benchmark-specific forgetting criteria. Beyond the complete SAUL framework, we further show on TOFU that applying the augmented-Lagrangian controller as a drop-in modifier to representative baselines improves their post-forgetting utility, demonstrating the practical value of explicit forgetting control.

BiasMix-Finance: Post-Generation KYC Guardrails for LLM Portfolio Advice

Large language models (LLMs) can generate plausible-sounding ETF portfolios while silently violating basic KYC-style constraints on risk, fees, and diversification. This is especially problematic in agentic multi-turn advisory systems, where each draft recommendation can become an action unless guarded by an auditable enforcement layer. We study a model-agnostic, asset-agnostic post-generation guardrail pipeline: (i) enforce a strict JSON allocation schema, (ii) validate allocations against numeric caps, and (iii) when violations occur, deterministically project the output to the nearest feasible portfolio via a convex quadratic program (QCQP). We introduce BiasMix-Finance (Mini), a compact stress-test benchmark for constrained decision-making under biased LLM generations, with a 16-ETF universe, three investor profiles, and eight bias prompts. Across three models and three inference modes (direct, critique, self-consistency), first-pass generations violate at least one cap in 47.6-85.7% of test cases (67.2% pooled), but the convex projection layer reduces final feasibility violations to 0% while requiring only a small correction distance (test pooled median D=||w*-w0||_2=0.066), indicating that the guardrail typically preserves the intent of the original allocation. We report violation rates and correction distances with confidence intervals, and paired model comparisons with multiple-testing correction. To support reproducibility, we release the dataset, prompts, caps, and code in our public GitHub repository.

Refusing Intent, Not Form: Wrapper-Based Intent-Group Supervision for LLM Safety

Safety tuning can improve harmful refusal, but models may learn surface-form shortcuts: wrapped harmful prompts bypass safety, while similarly wrapped benign prompts are over-refused. We propose Wrapper-Based Intent-Form Augmentation (WIFA), an automatic intent-group augmentation method that pairs wrapped harmful examples with structurally matched wrapped benign counterexamples, requiring no external teacher or manual per-wrapper intent labels. We use WIFA as a common data layer for two complementary fine-tuning routes: WIFA-Boost, a two-stage high-safety recipe, and Anchored Group-Consistent Refusal Training (A-GCRT), which regularizes refusal/compliance decision scores across same-intent wrappers and anchors harmful and benign groups on opposite sides of a margin. In the Qwen setting, WIFA-Boost reaches the strongest transformed-harmful refusal, while A-GCRT reduces OR-Bench over-refusal from 25.7\% for the base model to 17.4\%; reproduced baselines do not match these operating points. Llama results and ablations over data structure, two-stage order, and A-GCRT components support this intent-group interpretation without claiming universal below-base over-refusal.

Making Your LLMs More Objective: Stabilizing LLM Safety Behavior Across Traits with Trait-Invariant Safety Tuning

Aligned large language models (LLMs) are expected to exhibit safety behavior based on the content of the user request: they should refuse unsafe requests and comply with safe ones. However, we show that the same request can elicit substantially different safety decisions under different traits assigned in the system prompt, a failure mode we call trait-induced safety variation. To measure this failure, we introduce refusal-based metrics: Trait-Induced Deviation measures dataset-level deviation from the no-trait baseline, while Trait-Induced Flip Rate measures whether the same request receives different safety decisions across traits. We then provide a representation-level analysis of the mechanism behind trait-induced safety shifts and find that traits perturb the model's safety representations within a low-dimensional subspace. To achieve trait-invariant safety, where safety behavior remains stable across traits, we introduce Trait-Invariant Safety Tuning (TIST), a simple yet effective self-distillation framework that aligns an LLM's trait-conditioned behavior with its no-trait behavior. Guided by our analysis, we further propose Trait-Subspace Neutralization (TraSN), an instantiation of TIST, which enforces invariance only within the identified trait subspace. Experiments show that TraSN improves trait-invariant safety and strengthens harmful-request safety while preserving general capability. Our results highlight traits as an important factor in LLM safety and robust model behavior.

Learning to Persuade Exposes How Easily LLMs Abandon Correct Beliefs

Persuasion is a core dynamic of natural language communication, shaping how large language models (LLMs) update beliefs, resolve disagreements, and reach decisions. As LLMs increasingly debate, advise, and think collaboratively with humans and each other, resistance to harmful persuasion becomes a core requirement for reliable behavior. Yet we show that this requirement is far from met: a single targeted persuasive argument is enough to collapse model accuracy to near zero, even when the argument is factually false. We formalize this threat as adversarial persuasion and introduce an adversarial reinforcement learning framework that trains persuader agents to change a target model's answer in a single interaction. First, we show that optimizing persuasion strategies through trial and error exposes vulnerabilities that static prompting misses: RL-trained persuaders raise persuasion success from approximately 24% to over 93% against the training-time persuadee. Second, we find that these learned strategies transfer to unseen models, achieving 83% attack success on Qwen-14B, 79% on Llama-3.1-8B, and 25% on GPT-4o-mini. Third, we demonstrate that a curriculum that bootstraps on more persuadable open-weight models before targeting harder models further increases GPT-4o-mini attack success from 25% to 38%. Moreover, our results reveal that optimized persuaders increasingly rely on credibility-based tactics, including fabricated citations and false authoritative evidence. Together, these findings expose a critical weakness in current LLM agents: even when they initially reason correctly, they can be steered toward false conclusions by optimized natural language influence. This positions persuasion robustness as a necessary safety criterion for multi-agent and human-AI decision-making systems.

ProbGuard: Calibrated Safety Risk Estimation from LLM Output Distributions

Recent research on Large Language Model (LLM) safety has widely adopted guardrails to identify unsafe LLM outputs. Existing guardrails typically formulate safety assessment as a deterministic classification task, mapping a discrete token sequence to a discrete safety label. However, this paradigm has two limitations: First, safety assessment is inherently an uncertain problem, particularly during the early generation state. Second, relying solely on discrete token sequences discards the rich probabilistic information embedded in the LLM output distribution. To address these limitations, we propose the first completely probabilistic architecture-agnostic guardrail \textsc{ProbGuard} to leverage the LLM early output distributional signals for estimating and calibrating the safety probability, thereby enabling early stopping of unsafe ongoing outputs. Specifically, given an LLM's generated prefix distribution, we formulate the safety risk as the unsafe probability of its continued generation dynamics and estimate this risk by Monte-Carlo sampling. Through post-training on the distributional signals and calibrated safety risk, \textsc{ProbGuard} achieves the best calibration performance across all nine model--dataset combination settings, reducing the average Brier score and ECE by 79.6\% and 71.9\%, respectively, over the best baseline. \textsc{ProbGuard} further limits the attack success rate to at most 1\% across six representative jailbreak attacks after observing the LLM early output distributions from only the first ten decoding steps.

Toward a Theory of Value in AI Alignment

Can AI systems be aligned to human values? The popularization of large language models (LLMs) and multi-modal foundation models has seen a rise in harms spanning from toxic speech and hallucinations to AI agents executing unauthorized actions. Within the field of AI safety, these harmful instances are often framed as the alignment problem, or of models being misaligned with human values. Researchers have responded by pursuing applied and theoretical AI value alignment efforts, often without specifying what they mean by human values. How does the field of AI value alignment conceive of human values? How are these conceptions of values technically operationalized and evaluated? What does the emergent theory of value from this field signify for the future of AI? We annotated 94 value alignment research papers to discern their implicit theory of values in AI. The majority do not define values, relying heavily on preferences as a stand in that runs the risk of reducing complex culturally situated concepts down to binary choices. As researchers dispense with using human annotators for model training and evaluation, turning instead to synthetic data and autorater approaches to aligning and evaluating models, we identify the potential to close off alternative methods for contesting and enacting values in foundation models. In making AI value alignments philosophical commitments explicit, we seek to bring great specificity and under explored perspectives in the debate on whether and how AI can address human values.

Safety Cost of Steering Vectors Is Separable and Reducible

Steering vectors are a lightweight tool for controlling LLM behavior. However, emerging evidence shows that steering vectors can unintentionally compromise a model's safety mechanisms and increase compliance with harmful requests, while no effective mitigation yet exists. In this work, we show that this safety degradation arises from a separable component in the vector that disrupts the model's safety mechanisms but contributes little to the steering objective. We identify and remove this safety-degrading component, formulating the task as a constrained optimization problem solved through primal-dual updates, subject to preserving the intended steering effect and bounding false refusal. The resulting solution is both interpretable and surgical: the optimization recovers a single direction whose ablation from the steering vector restores model safety with minimal utility cost. Across models, steering behaviors, and attack suites, including unseen attacks types, our method substantially reduces steering-induced safety degradation while preserving the original steering effect with minimal impact on false refusal. Our method offers a post-hoc correction to steering vectors that mitigates their safety cost, and more broadly, it provides a general recipe for applying activation-level model interventions without paying a safety tax.

GROM: Gradient-Free Rapid One-Shot Machine Unlearning

Machine unlearning has become a critical capability for safely removing specific, sensitive knowledge from large language models (LLMs). Current state-of-the-art approaches primarily rely on iterative, training-time unlearning via fine-tuning. However, even when utilizing parameter-efficient dimensionality reduction techniques like LoRA, gradient-based optimization remains computationally expensive and lacks explicit analytical formulations. It can also leave the targeted knowledge merely hidden rather than removed, to the point that simply quantizing the unlearned model restores much of what it was supposed to have erased. To resolve this, we propose a novel one-shot unlearning approach, abandoning iterative optimization in favor of a direct, exact analytical solution. We frame the unlearning process as a ridge-regularized least-squares optimization problem, deriving a closed-form additive update for targeted weight matrices. This update forces the selected layer to suppress unwanted content while strictly preserving its behavior on retained data. Computed from gradient-free forward passes alone, with no backpropagation and no iteration to convergence, GROM applies the weight edit in mere seconds, which makes it orders of magnitude faster than traditional fine-tuning. Extensive evaluations demonstrate that GROM achieves state-of-the-art forgetting-utility trade-offs on TOFU-5%, TOFU-10%, MUSE-Books, MUSE-News and WMDP, significantly reducing computational overhead without sacrificing overall model performance. Because the update removes the targeted content from the weights instead of masking it, GROM also withstands the low-bit quantization attack that recovers much of the content a gradient-based baseline had appeared to forget. Our code is publicly available at https://github.com/Batorskq/GROM.