Safety & alignment

Backdoor attacks on language models pose a significant threat to AI safety, where models behave normally on most inputs but exhibit harmful behavior when triggered by specific patterns. Detecting such backdoors through mechanistic interpretability remains an open challenge. We investigate two sparse autoencoder architectures -- Crosscoders and Differential SAEs (Diff-SAE) -- for isolating backdoor-related features in fine-tuned models. Using a controlled SQL injection backdoor triggered by year-

Recent advancements in visual context compression enable MLLMs to process ultra-long contexts efficiently by rendering text into images. However, we identify a critical vulnerability inherent to this paradigm: lowering image resolution inadvertently catalyzes jailbreaking. Our experiments reveal that the safety defenses of SOTA models deteriorate sharply as resolution degrades, surprisingly persisting even when text remains legible. We attribute this to ``Cognitive Overload'', hypothesizing that

AI agents that plan, retain memory across sessions, invoke external tools and act with partial autonomy are transforming human--AI collaboration. Research on affective computing, simulated empathy in large language models, trust in automation and AI safety has illuminated important design principles, yet these literatures remain fragmented. No integrated account explains how affective cues operate within agentic collaboration -- settings in which humans delegate, monitor and correct consequentia

Large language models often require fine-tuning to better align their behavior with user intent at deployment. Existing approaches are commonly divided into online and offline paradigms. Online methods, such as RL-based alignment, can directly optimize outcome quality but typically rely on external reward models and iterative rollouts, making them costly and difficult to deploy in many cases. Offline methods are more efficient, but prevailing approaches such as supervised fine-tuning (SFT) and d

Language model (LM) alignment improves model outputs to reflect human preferences while preserving the capabilities of the base model. The most common alignment approaches are (i) reinforcement learning, which maximizes the expected reward under a KL-divergence constraint, and (ii) best-of-$N$ alignment, which selects the highest-reward output among $N$ independent samples. Despite their widespread use, the fundamental limits of reward improvement under a KL budget remain poorly understood. We c

Causal abstraction offers a principled framework for mechanistic interpretability, aligning a high-level causal model with the low-level computation realized by a neural network through counterfactual intervention analysis. Existing methods such as distributed alignment search (DAS) learn expressive subspace interventions, but the relevant neural site is unknown a priori, so finding a handle requires a computationally burdensome search over candidate sites. We introduce PLOT (Progressive Localiz

Adaptive test-time compute for LLM agents aims to invoke extra computation only when it improves performance. Existing methods typically use confidence-, uncertainty-, or difficulty-based gates, assuming a fixed direction from the gating signal through compute need to the value of computation. This makes gating a utility-calibration problem: gating signals should align with whether extra computation improves the final outcome over the base policy. We show that this alignment is unstable: the sam

Diffusion language models (DLMs) have recently demonstrated capabilities that complement standard autoregressive (AR) models, particularly in non-sequential generation and bidirectional editing. Although recent work has shown that pretrained autoregressive checkpoints can be converted into diffusion language models, existing recipes primarily transfer parameters through continued denoising training with objective- and attention-level modifications. We instead ask whether the internal representat

Reinforcement Learning (RL) has emerged as a crucial paradigm for unlocking the advanced reasoning capabilities of Large Language Models (LLMs), encompassing frameworks like RLHF and RLAIF. Regardless of the specific optimization algorithm (e.g., PPO, GRPO, or Online DPO), online RL inherently requires an exploratory trajectory generation (rollout) phase. However, for long-context reasoning tasks, this rollout phase imposes a severe ``memory wall'' due to the exorbitant Key-Value (KV) cache foot

Graph Edit Distance (GED) is a fundamental, albeit NP-hard, metric for structural graph similarity. Recent neural graph matching architectures approximate GED by first encoding graphs with a Graph Neural Network (GNN) and then applying either a graph-level regression head or a matching-based alignment module. Despite substantial architectural progress, the role of encoder geometry in neural GED estimation remains poorly understood. In this paper, we develop a theoretical framework that connects

We propose a deterministic adjoint matching framework that formulates human preference alignment for flow-based generative models as an optimal control problem over velocity fields. One can directly regress the control toward a value-gradient-induced target under the current policy, leading to a simple and stable training objective. Building on this perspective, we introduce a truncated adjoint scheme that focuses computation on the terminal portion of the trajectory, where reward-relevant signa

We identify and formalize an underexplored phenomenon in deep learning optimization: directional alignment and loss convergence can be decoupled. An optimizer can exhibit near-perfect directional consistency (cc_t -> 1, measured via consecutive gradient cosine similarity) while the loss remains high or decreases slowly. This observation reveals that existing optimizers such as Adam, SGD, and RMSprop lack explicit mechanisms to exploit temporal consistency in gradient directions, relying instead

Recent advances in diffusion transformers (DiTs) have enabled promising single-turn image editing capabilities. However, multi-turn editing often leads to progressive semantic drift and quality degradation.In this work, we study this problem from a latent-space frequency perspective by decomposing the editing process into two functional components: VAE and DiT. Through systematic analysis in the VAE latent space, we uncover that the DiT introduces dominant low-frequency drift that accumulates as

Outcome metrics can certify the wrong behavior. We study this failure in a two-hotel revenue-management simulator where Hotel A trains an agent against a fixed rule-based revenue-management competitor, Hotel B. A standard learning agent can obtain near-reference revenue per available room (RevPAR) while failing to learn market-like yield management: it sells too aggressively, undercuts, or collapses to modal price buckets. We diagnose this as a Goodhart-style failure under partial observability.

A leading proposal for aligning artificial superintelligence (ASI) is to use AI agents to automate an increasing fraction of alignment research as capabilities improve. We argue that, even when research agents are not scheming to deliberately sabotage alignment work, this plan could produce compelling but catastrophically misleading safety assessments resulting in the unintentional deployment of misaligned AI. This could happen because alignment research involves many hard-to-supervise fuzzy tas

Large language model (LLM) alignment via reinforcement learning from human preferences (RLHF) suffers from unstable policy updates, ambiguous gradient directions, poor interpretability, and high gradient variance in mainstream pairwise preference learning paradigms. To systematically address these limitations, we establish a unified theoretical framework for preference-based RL optimization centered on the Pair-GRPO family, comprising two tightly coupled variants: Soft-Pair-GRPO and Hard-Pair-GR

Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation. We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request, and present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase

Large Language Models (LLMs) are evolving into autonomous trading agents, yet existing benchmarks often overlook the interplay between architectural reasoning and strategy consistency. We propose Strat-LLM, a framework grounded in Stratified Strategy Alignment. Operating in a live-forward setting throughout 2025, it integrates heterogeneous data including sequential prices, real-time news, and annual reports to eliminate look-ahead bias. Extensive stress tests on A-share and U.S. markets reveal:

The safety alignment of Large Language Models (LLMs) remains vulnerable to Harmful Fine-tuning (HFT). While existing defenses impose constraints on parameters, gradients, or internal representations, we observe that they can be effectively circumvented under persistent HFT. Our analysis traces this failure to the inherent redundancy of the high-dimensional parameter space: attackers exploit optimization trajectories that are orthogonal to defense constraints to restore harmful capabilities while

Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input. We investigate the root causes of these failure modes with a mechanistic analysis focusing on the decoder-based VLMs. We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with

Heterogeneous federated learning (HtFL) aims to enable collaboration among clients that differ in both data distributions and model architectures. Prototype-based methods, which communicate class-level feature centers (prototypes) instead of full model parameters, have recently shown strong potential for HtFL. Existing prototype-based HtFL methods typically reuse the MSE-based or cosine-based alignment mechanism developed for homogeneous FL when aligning client-specific representations with glob

Federated learning (FL) systems typically employ stateless client selection, treating each communication round independently and ignoring accumulated evidence of client contribution quality. Under non-IID data, this leads to slow convergence and unstable training, particularly when selection relies on local proxies (e.g., training loss) that are misaligned with the global optimization objective. These challenges are especially pronounced in Internet of Things (IoT) and Industrial IoT (IIoT) envi

Large reasoning models (LRMs) sometimes note in their chain of thought (CoT) that they may be under evaluation. Researchers worry that this verbalised evaluation awareness (VEA) causes models to adapt their outputs strategically, optimising for perceived evaluation criteria, which, for instance, can make models appear safer than they actually are. However, whether VEA actually has this effect is largely unknown. We tested this across open-weight LRMs and benchmarks covering safety, alignment, mo

Posterior collapse in variational autoencoders is often diagnosed by its symptoms: a small KL term, a strong decoder, or weak use of the latent code. These signals are useful, but they do not define a collapse boundary. We study a concrete failure mode, input-independent constant collapse, and show that this case admits an exact threshold. For any fixed nonconstant teacher distribution \(T(\cdot\mid x)\), the best constant student is the dataset-average teacher distribution, and its alignment co

Intent-obfuscation-based jailbreak attacks on multimodal large language models (MLLMs) transform a harmful query into a concealed multimodal input to bypass safety mechanisms. We show that such attacks are governed by a \emph{reconstruction--concealment tradeoff}: the transformed input must hide harmful intent from safety filters while remaining recoverable enough for the victim model to reconstruct the original request. Through a reconstruction analysis of three representative black-box methods

Estimating individualized treatment effects from longitudinal observational data is central to data-driven medicine, yet existing methods face a fundamental limitation: reducing confounding bias often suppresses clinically informative heterogeneity, degrading patient-specific predictions. Here, we identify this tension as a bias-precision paradox in causal representation learning and introduce sampling-based maximum mean discrepancy (sMMD), a stochastic alignment strategy that replaces global ad

Recent developments in AI safety research have called for red-teaming methods that effectively surface potential risks posed by generative AI models, with growing emphasis on how red-teamers' backgrounds and perspectives shape their strategies and the risks they uncover. While automated red-teaming approaches promise to complement human red-teaming through larger-scale exploration, existing automated approaches do not account for human identities and rarely incorporate human inputs. In this work

Vector search systems built on frozen vision encoders face queries from unseen classes at deployment, yet existing adapter training collapses under this shift: high-capacity adapters with global contrastive losses silently reassign unseen-class samples to wrong seen-class clusters, dropping worst-case Label Precision by over 40 points below the frozen baseline in our tests. We propose Euclidean Geodesic Alignment (EGA), a residual adapter that couples three principles: zero initialization, local

Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within

Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced guardrails remain vulnerable to such attacks despite advances in safety alignment and external guardrails. In this work, we address this challenge by detecting t

In this paper, we present X-Voice, a 0.4B multilingual zero-shot voice cloning model that clones arbitrary voices and enables everyone to speak 30 languages. X-Voice is trained on a 420K-hour multilingual corpus using the International Phonetic Alphabet (IPA) as a unified representation. To eliminate the reliance on prompt text without complex preprocessing like forced alignment, we design a two-stage training paradigm. In Stage 1, we establish X-Voice$_{\text{s1}}$ through standard conditional

Causal representation learning (CRL) seeks to recover latent variables with identifiability guarantees, typically up to permutation and component-wise reparameterization under appropriate assumptions. However, identifiability does not imply interpretability: latent semantics are typically assigned post hoc by alignment with known ground-truth factors. This limitation is particularly acute in scientific time series, where underlying mechanisms are unknown and discovering interpretable structure i

Large language models (LLMs) remain vulnerable to adversarial prompting despite advances in alignment and safety, often exhibiting harmful behaviors under novel attack strategies. While adversarial training can improve robustness, existing approaches are computationally expensive and difficult to scale. Recent continuous adversarial training methods, such as Continuous adversarial training (CAT) and Continuous Adversarial Preference Optimization (CAPO), address this challenge by leveraging gradi

This position paper argues that sycophancy in LLMs is a boundary failure between social alignment and epistemic integrity. Existing work often operationalizes sycophancy through external behavior such as agreement with incorrect user beliefs, position reversals, or deviation from an objective standard of correctness. These formulations capture only overt forms of the phenomenon and leave subtler boundary failures involving epistemic integrity and social alignment underspecified. We argue that sy

The rapid adoption of AI coding agents has produced a dominant workflow pattern -- often called "vibe coding" -- that prioritizes speed of implementation over deliberate preparation. We argue that this approach creates a systematic alignment problem: agents that lack sufficient context produce code requiring extensive debugging and refactoring, consuming substantial development time. Drawing on the culinary concept of mise en place (everything in its place; abbreviated MEP), we propose a three-p

Millimeter-wave (mmWave) frequencies promise multi-gigabit connectivity for vehicle-to-everything (V2X) networks, but face challenges in terms of severe path loss and mobility-related beam misalignment. Reliable V2X connectivity requires fast, double-directional beam alignment. However, existing methods suffer from high training overhead and limited generalization to unseen scenarios. This paper presents VIsion-based BEamforming(VIBE), a hybrid model-based, closed-loop, learning architecture for

Recent flow matching (FM) methods improve the few-shot adaptation of vision-language models, by modeling cross-modal alignment as a continuous multi-step flow. In this paper, we argue that existing FM methods are inherently constrained by incompatible geometric priors on pre-trained cross-modal features, resulting in suboptimal adaptation performance. We first analyze these methods from a polar decomposition perspective (i.e., radial and angular sub-manifolds). Under this new geometric view, we

Reward models are a key component of large language model alignment, serving as proxies for human preferences during training. However, existing evaluations focus primarily on broad instruction-following benchmarks, providing limited insight into whether these models capture socially desirable preferences. As a result, important failures in social alignment can remain hidden. We extend reward-model benchmarking to four socially consequential domains: bias, safety, morality, and ethical reasoning

As Large Language Models (LLMs) evolve into interactive agents, understanding their behavioral alignment within human social dynamics becomes essential. While behavioral game theory offers a framework to study these interactions, previous work has predominantly relied on outcome-based metrics such as average scores. This focus overlooks the mechanisms that facilitate sustainable cooperation, as identical scores can be derived from vastly different strategies. To bridge this gap, we introduce SOD

AI-assisted development tools enable rapid prototyping of services but often lack awareness of architectural constraints, infrastructure dependencies, and organizational standards required in production environments. Consequently, generated artifacts may exhibit brittle behavior and limited deployability. We propose a retrieval-augmented scaffolding approach that combines platform-based code generation with agentic clarification loops to expose and resolve architectural constraint ambiguities. B