Value Alignment

As language-model-based AI is increasingly deployed in autonomous settings, aligning its goals and values with those of humans becomes critical. Today, alignment, and the assistant identity itself, are typically introduced only after pretraining, once behavioral priors are already established. This can make values a thin overlay, rather than deeply rooted, and facilitate subsequent misalignment. Pursuing a different paradigm, we introduce Synthetic Persona Pretraining (SPP), which installs the desired assistant persona from token zero in pretraining. First, we annotate pretraining documents with value-aligned first-person reflections derived from a normative value constitution. Second, we pretrain via the standard cross-entropy loss on standard pretraining documents as well as their reflections, which installs the desired persona among a multitude of other personas. Finally, we post-train on user-assistant dialogue data, which binds this desired persona to the assistant identity, a process we call persona binding. By pretraining models up to 3B parameters on 500B tokens, we show that SPP improves constitution following and jailbreak robustness, and reduces the misalignment rate in out-of-distribution moral dilemmas, while preserving capabilities. Early intervention matters: compared with alignment from token zero, introducing SPP only at the end of pretraining yields weaker constitution adherence, does not shift value priorities, and leads to less aligned choices in dilemmas. This advantage depends on persona binding and, importantly, increases with pretraining budget. Overall, our results show that shaping values early is critical for alignment and establish pretraining-time persona interventions as an effective approach to do so.

Hybrid-thinking multimodal large language models (MLLMs) allow a single model to alternate between deliberative thinking and latency-efficient non-thinking inference. Although these modes differ in reasoning budget, their delivered responses should satisfy the same user-facing standard. Correctness alone may not characterize this response quality; we therefore evaluate task accuracy and response-pattern failures as complementary outcomes. We study this gap through \textbf{response-pattern alignment}: whether thinking and non-thinking interfaces preserve acceptable final-response behavior. We introduce \textbf{PatternEval}, a failure-enriched diagnostic benchmark comprising 2,415 multimodal prompts spanning visual perception and grounding, structured image understanding, and multimodal knowledge reasoning. PatternEval tests four recurrent failures: chain-of-thought leakage, response repetition, logical contradiction, and performative reasoning. Response-pattern failures are widespread across models from different providers, with non-thinking inference exhibiting substantially higher failure rates and thereby creating systematic misalignment between thinking and non-thinking interfaces. Motivated by this diagnosis, we develop \textbf{PatternRM}, a response-level reward model, and \textbf{PatternRL}, which introduces pattern-specific penalties during reinforcement learning. Experiments on Qwen3-VL-4B and Qwen3-VL-8B show that incorporating pattern-specific penalties into reinforcement learning can mitigate cross-mode misalignment while incurring a marginal task performance trade-off. Together, PatternEval and PatternRL provide an evaluation-and-training framework for aligning user-visible response patterns across hybrid-thinking interfaces.

State-aligned distortion has been documented in China-origin text-based large language models (LLMs), but whether, and in what form, it arises in multimodal systems has not been systematically examined. We construct a balanced benchmark of 200 core entries spanning ten politically sensitive topics, plus a seven-variant visual-abstraction probe, and run nine vision-language models (VLMs), seven China-origin and two non-China, across four elicitation paradigms and two prompt languages, yielding 21,708 trials. Each response is audited on six dimensions -- explicit refusal, information integrity, visual grounding, state-aligned framing, language consistency, and response length -- by two independent frontier LLM judges, validated against three human experts on a 200-trial sample. Measuring each dimension separately lets us decompose multimodal censorship into individual signals rather than a single refusal-based score; in particular, refusal and framing are measured independently, so a model can stop refusing while still reframing. We find that (i) Chinese-language prompting roughly triples the odds of state-aligned framing, within every model; (ii) China-origin models reframe more than non-China models (direction robust across judges and human raters; magnitude 1.6--3.2x); (iii) the effect is strongest in text-only political commentary (36.5%) and is gated by recognition of the depicted subject rather than pixel detail, persisting even at silhouette for iconic images; and (iv) across four Qwen generations, state-aligned framing rises while explicit refusal falls: censorship migrates from a visible act (refusal) to an invisible one (fluent reframing). We argue this shift to invisible reframing is fundamentally a problem of human-AI interaction: it removes the very signal users rely on to recognize that information has been withheld.

Safety alignment in large language models is often treated as a distributed property of the entire network, yet its practical brittleness suggests that refusal behavior may be concentrated in a smaller set of parameters. This work addresses where safety-aligned refusal is encoded by transplanting weights from aligned models into matched unaligned base models at multiple levels of granularity. Using two open-weight model pairs and four safety benchmarks, we conducted experiments to compare the effects of replacing attention weights, MLP weights, contiguous layer regions, and MLP blocks. Across both model families, refusal transfer is dominated by MLP weights: replacing MLP parameters recovers substantially more malicious-prompt refusal than replacing attention parameters, with gains of at least 2.7 times more across benchmarks. Within the MLP stack, refusal-relevant parameters exhibit a consistent mid-network concentration, as the block spanning layers 8-11 is selected first in all six greedy searches over model-dataset pairs. The results also show that the composition of safety-relevant components is non-additive: in five of six greedy trajectories, adding more aligned blocks can reduce refusal performance, and selective block subsets can outperform full MLP transplantation on malicious refusal, benign over-refusal, or both. Finally, greedy orders transferred to OR-Bench vary with the source benchmark used to derive them, indicating a benchmark-dependent precision-coverage trade-off. These results suggest that safety alignment in current LLMs is both localized and interaction-sensitive, offering insight into alignment brittleness and potential avenues for targeted safety interventions.

Group alignment adapts a language model to a demographic group to produce responses that reflect the group's opinions, values, and preferences. Sycophancy, a well-documented by-product of alignment, causes the model to over-agree with the user regardless of factual and objective information. However, existing group alignment methods and evaluations focus only on how closely the model matches the group's opinions, overlooking the induced change in sycophantic behaviour. To bridge this gap, we introduce \textbf{G}roup \textbf{A}lignment-induced \textbf{S}ycophancy (GAS) and systematically evaluate alignment across 3 methods, 4 models and 13 demographic groups, on both the intended gain in opinion alignment and the unintended shift in sycophancy. We find that gain and shift are non-uniform across groups: under an identical budget, some groups receive larger gains in opinion alignment than others, and the induced sycophancy shift forms a group-specific profile rather than a single-dimensional change. These results suggest that group alignment should be reported as a two-sided, multi-dimensional profile rather than a single fit score that accounts for per-group differences when adapting LLMs to diverse populations.

Traditional offline recommendation evaluation relies heavily on complex, manually maintained feature pipelines that are difficult to scale. While Large Language Models (LLMs) offer a promising alternative by predicting user engagement directly from raw text logs, empirical analysis in this study identifies a critical failure mode termed bidirectional rationalization. In a zero-shot setting, LLMs are found to convincingly argue for both positive and negative user engagement outcomes on the exact same item with identical evidence, highlighting the unreliability of off-the-shelf LLMs in predicting user engagement. To resolve this, we develop and apply a sequential behavioral alignment framework pairing fine-tuning with preference optimization over paired correct and counterfactual rationales. Evaluated on real-world homepage interaction logs, this aligned reasoning approach achieves a 32.19% lift in Macro-F1 score over the zero-shot baseline and matches the production feature-engineered baseline. The results demonstrate that behavioral alignment mitigates bidirectional rationalization while delivering human-interpretable reasoning traces without manual pipeline overhead.

AlphaFold3-class structure predictors pair a representational trunk, which processes sequence and context, with a diffusion module, which generates atomic coordinates. How biological information changes as it crosses this architectural boundary remains poorly understood. We analyze per-residue activations from the Pairformer trunk and diffusion module of Boltz-1 using linear probes, sparse autoencoders (SAEs), and causal interventions. From the trunk, both geometry (secondary structure, disorder) and sequence chemistry (amino-acid identity, signal peptides, disulfide-bond annotations) are linearly decodable. In the diffusion module, the two diverge. Secondary structure transfers essentially unchanged, whereas sequence chemistry is strongly attenuated. We then test whether decodable directions can steer the model, intervening on the final trunk single representation that conditions the diffusion module. Helix and coil directions change predicted structure dose-dependently against matched-norm random controls, but a beta-strand direction that is highly predictive (F1 =0.82) produces no measurable increase in strand content: linear decodability does not imply causal influence at the site we tested. The same probes also score markedly lower against sparse SwissProt annotations than against dense DSSP labels, because unannotated residues that the model gets right are charged as false positives; such scores are therefore lower bounds. Finally, supervised probes outscore single SAE features wherever a label already exists. We release the trained trunk and diffusion SAEs, Boltz-1 per-residue activations, and the analysis code.

Graph learning presupposes a graph, and tables and relational databases do not come with one. Applying a GNN to them requires deciding which entities become nodes, which of them to connect, and through which relations---a decision made by hand, by schema heuristics, or by training a model on every candidate graph and keeping the best. We give a criterion that requires no trained graph model. In the minimal table-to-graph abstraction each row is a node, so a message-passing GNN, bounded by 1-WL, sees a construction only as a partition of the rows into colour-refinement classes: a construction is good for a task when that partition separates rows with different labels and does not split rows that share one. AutoGrable turns this criterion into a construction procedure. For incidence constructions the partition is fixed by the selected columns, so building a graph reduces to choosing them, and we score a candidate subset by a label-alignment risk: the held-out risk of the best predictor constant on its blocks, penalised by an occupancy term measuring how thinly the blocks are populated. The score materialises no graph and trains no GNN, so AutoGrable can search the space of subsets greedily and cheaply, and returns the resulting grable for single tables and for foreign-key schemas alike. Our experiments show that over a space of candidate graphs the score discards a large fraction while retaining the best; that AutoGrable recovers the columns that generate the label on controlled tasks and outperforms fixed, random, and task-aware constructors on real tasks under a fixed predictor; and that it is the only method compared that can decline to build a graph when none helps.

The lack of diversity in LM content is widely attributed to the alignment process, but how and where exactly in the pipeline this collapse begins is unknown. We argue that output homogeneity is likely learned during the pretraining phase, and only \emph{revealed} or magnified during the alignment process. Specifically, we find that semantic convergence is observed from the first alignment stage--the instruction-tuning phase (SFT)--suggesting that homogeneity might already exist in the pre-alignment model. To investigate this, we conduct controlled SFT experiments examining how training data influences output convergence on specific input/output pairs. We find that convergence can be revealed and amplified, but not introduced by the SFT data, supporting its role as a catalyst rather than a cause. To further test whether homogeneity originates before alignment, we measure convergence in base models. We find that instruct-like collapse can be induced through prompting alone, even without alignment. Taken together, our results suggest that semantic convergence may arise naturally from the objectives underlying LM training, making it difficult to mitigate through post-alignment interventions alone.

The Workshop on Trustworthy Natural Language Processing (TrustNLP), co-located with major ACL conferences since 2021, has grown from 8 proceedings papers to 41 over six editions, documenting a field-wide transition from post-hoc interpretability of static models to mechanistic understanding and proactive control of generative systems. We synthesize insights from all 144 proceedings papers, classifying them along six trust dimensions grounded in established frameworks (TrustLLM, DecodingTrust). We observe co-occurrences with capability emergence. The release of the first high-impact chat models activated all trust dimensions simultaneously, while subsequent model generations shifted focus toward truthfulness and safety alignment. Analysis from the classification study reveals that truthfulness is the fastest-growing dimension (absent in 2021-2022, comprising 37% of papers by 2025-2026), fairness remains the most consistent theme, and explainability exhibits a U-shaped trajectory; declining as post-hoc methods lost relevance but resurging in 2026 through mechanistic interpretability. A cross-venue comparison with ACL, NAACL, EACL, and EMNLP (~2K papers) in the same period shows that TrustNLP's topical distribution closely follows the field average. We identify four structural insights and conclude with actionable directions for the research community.

Existing Multimodal Large Language Models (MLLMs) predominantly rely on image-text pairs for modality alignment pretraining, mapping global image representations to long textual descriptions. However, this image-level alignment suffers from referential ambiguity: models struggle to infer the correspondences between multiple visual objects and textual entities from the global representation, leading to data inefficiency and suboptimal semantic grounding. To address this, we propose MultiModal Code-Switching (MMCS), a novel pretraining paradigm that provides explicit object-level supervision. Inspired by the linguistic phenomenon of code-switching, MMCS interleaves vision and language by replacing textual entities with their corresponding visual objects, enforcing local vision-language grounding. We further develop a scalable data synthesis pipeline to generate a pretraining dataset of 773K samples with accurate object-entity correspondences. Experiments show that MMCS is highly data-efficient: with only 50K samples, it matches or surpasses models trained on 600K image-text pairs. Furthermore, MMCS consistently improves visual grounding and perception capabilities across varying model scales.

Emergent misalignment (EM) is the phenomenon where fine-tuning a language model on a narrow task leads to harmful behavior in unrelated domains. A leading mechanistic account attributes EM to persona features: latent directions acquired during pre-training that misaligned fine-tuning amplifies. We ask where these features come from: which pre-training documents activate them, and whether naturally occurring human-written text suffices to induce EM. Using Sparse Autoencoder (SAE) based model diffing across four open-weight models, we find that features related to jailbreak personas, sarcasm, deception, and manipulation are amplified by misalignment fine-tuning, while safety-relevant and assistant-identity features are suppressed. Steering individual features controls EM in both directions: it induces misalignment rates of up to 62% in aligned models -- exceeding the 35% reached by misalignment fine-tuning itself -- and re-aligns misaligned models to near-baseline misalignment rates. Attributing the causal features to a corpus of one million pre-training web documents retrieves semantically relevant narratives about villainous characters, domination, and harmful agency. However, fine-tuning on these human-written documents does not reliably induce EM, even after reformatting into assistant-style responses, whereas synthetic instruction-response pairs derived from the same content do -- and transfer across model families. Semantic relevance alone is therefore not sufficient: response structure or model-generated phrasing plays an important role in inducing EM.

Understanding both expressed and evoked emotions is critical for multimodal large language models (MLLMs) to achieve comprehensive affect-aware interactions. However, existing benchmarks typically examine expressed and evoked emotions in isolation or are constrained to coarse-grained and incomplete affective characterizations. To bridge this gap, we introduce E$^3$mo-Bench, a scalable benchmark comprising $12{,}314$ question-answer pairs across $2{,}524$ videos with predefined affective perspectives. It evaluates evoked and expressed emotion understanding via $3$ complementary tasks: emotion perception, open-vocabulary recognition, and valence-arousal-dominance (VAD) assessment. To efficiently scale reliable continuous annotations, we propose Bayesian Pairwise Alignment, which aggregates sparse, low-burden pairwise judgments into anchor-referenced VAD estimates. Furthermore, we develop E$^3$mo-Score, a training-free agent that aggregates complementary judgments from a five-model committee to improve VAD estimation. Extensive experiments validate the effectiveness of our framework and expose a pronounced performance skew between evoked and expressed emotion paradigms. These findings, coupled with MLLMs' persistent deficits in fine-grained recognition and dimensional assessment, chart a clear course for advancing multimodal emotional intelligence.

Large language models (LLMs) increasingly act in interactive settings where their behavioral styles affect user experience, safety, and downstream decision making. Existing LLM personality studies largely rely on self-report questionnaires administered in first-person settings, making the resulting profiles sensitive to surface elicitation choices and poorly grounded in concrete model behavior. In this work, we introduce a situated behavioral-data (B-data) framework for studying and controlling LLM behavioral personality. We construct 3,200 contrastive behavioral scenarios spanning 20 behavioral patterns and four prompt registers, grounded in validated psychometric facets such as BFI-2, DOSPERT, and HEXACO. Using this framework, we find that LLMs exhibit stable and model-specific behavioral profiles, while also revealing register-dependent shifts across first-person decisions, advice-giving, and task execution. We then show that these behavioral patterns can be controlled through Behavioral Mode Axes (BMAs), activation-space directions derived from contrastive behavioral traces. Compared with response-derived BMAs, which are more prone to trait drift, thought-derived BMAs more faithfully capture the intended behavioral mechanism and provide cleaner control over situated behavioral styles. Our results suggest that LLM personality-like tendencies are better understood not as abstract self-report traits, but as measurable and controllable behavioral modes grounded in concrete interaction contexts. Our code and data are available at https://github.com/lhz191/LLM-Behavioral-Personality.

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.

Long speech separation typically employs a segment-separation-stitch paradigm where recordings are divided into short segments, processed independently, and stitched together. Its challenge lies in predicting cross-segment permutations. This paper proposes a training-free dynamic clustering approach for cross-segment permutation alignment using speaker embedding reference pools. The method predicts the permutation using the cosine similarity between current segment embeddings and the reference pools. The approach updates reference pools by retaining the most representative speaker embeddings based on their overall cosine similarity with existing references. As a plug-and-play post-processing module compatible with existing separation models, the proposed method demonstrates superior performance compared to existing methods on dense and sparse long speech scenarios, particularly in challenging sparse scenarios with extended utterance gaps, and further shows robustness to speaker count estimation errors in unknown speaker count scenarios.

Web navigation agents are capable of addressing various types of tasks on different websites. Current baselines on web navigation are either unimodal or lack strong reasoning abilities given multimodal inputs. Focusing on the WebShop benchmark, a real-world website simulation, we explore the alignment of text and images, as well as multimodal reasoning and planning abilities, to enhance the performance of web navigation agents. We propose three innovative multimodal enhancements: Multimodal Enhanced LLM for Online Navigation (MELLON), VQAgent, and Multimodal Ranker. MELLON demonstrates a significant improvement in task completion accuracy, with a 9.26% increase after just one epoch of training. Our findings suggest the necessity of further exploration into multimodal approaches, with a focus on more extensive training and alignment strategies to enhance the effectiveness of web navigation agents.

Image matting is an essential enabling technology for modern visual content production, where foreground extraction determines the realism and editability of downstream creation workflows. However, precise alpha estimation in open-world scenes remains challenging because real foregrounds exhibit highly diverse appearances and opacity patterns. This makes existing methods struggle with semantic ambiguity and fine-grained opacity variation, especially in sparse boundary regions that are fragile and difficult to supervise. To address this gap, we present RenderMatte, a trimap-guided matting framework that adapts FLUX.1 Kontext through full-parameter fine-tuning, leveraging image editing priors for structure-preserving alpha prediction. During supervised adaptation, an alpha-edge objective preserves the latent flow-matching signal while strengthening pixel-space boundary supervision. We further introduce group-relative alpha alignment for post-training. It compares multiple mattes sampled under the same trimap condition using matting-specific rewards for alpha accuracy, boundary fidelity, trimap compliance, and compositional consistency. To overcome the lack of precise edge annotations, we construct the RenderMatte dataset, a large-scale synthetic dataset combining 3D-rendered RGBA foregrounds with diverse multi-source assets. It features exact strand-level alpha annotations and diverse background composites. Experiments show state-of-the-art performance across all benchmarks, demonstrating a scalable path toward high-fidelity matting in open-world scenes.

In-context learning (ICL) can induce emergent misalignment (EM), where narrow misaligned examples alter answers to unrelated questions. Existing prompts, however, conflate harmful-text exposure with an invitation to continue assistant behavior. We hold harmful answers fixed while varying their delivery as demonstrations, evidence, assistant history, or tool output. Across ten independently sampled contexts, demonstration framing raises broad EM by $30$--$32$ percentage points on a susceptible Gemini model; the gap survives domain exclusion, semantic clustering, unseen questions, and four prompt templates. Format and length-matched controls show that harmful content is necessary but insufficient. A role times continuation factorial further reveals model-dependent provenance effects: Gemini follows both assistant and tool histories, whereas Grok largely resists tool-framed continuation. Several other frontier and open-weight models show no gap. Blinded human audits confirm every main contrast and show that the model judge underestimates active-condition failures. Thus continuation framing is a strong, model-dependent moderator of ICL-EM, not a universal consequence of harmful context.

Attention mechanisms have been widely utilized in modern deep learning, and many existing multi-omics models inherit their conventional use to allow unrestricted bidirectional interactions. However, the fundamental logic of life is directional. Existing designs often overlook the directionality suggested by the central dogma, potentially limiting transfer across heterogeneous cancers, downstream tasks, and incomplete modality settings. In this work, we present DoGMA, a central-dogma-guided foundation model for pan-cancer multi-omics analysis, arguing that robust transfer requires representations with domain-specific inductive bias. Concretely, we build it on a Transformer-MoE architecture where directed attention biases inter-omics communication toward central-dogma information flow. We further pretrain our model with masked hierarchical omics reconstruction to guide it toward learning central-dogma-consistent interactions. Across diverse downstream tasks, including cancer representation learning, survival prediction, and metastasis prediction, DoGMA consistently demonstrates strong predictive performance. Ablations and analyses further suggest that the performance gains arise from the synergy between central-dogma-guided directed attention and reconstruction-based pretraining, which together promote more biologically consistent cross-omics information exchange. Overall, DoGMA demonstrates that domain-specific inductive biases can improve the robustness and transferability of multi-omics foundation models, offering new insights into the design of attention mechanisms for multi-omics representation learning.

Aligning large language models (LLMs) with human values remains a major challenge, especially for trustworthy AI. While existing approaches such as RLHF, CAI, and their variants have achieved promising results, they often rely on a single-agent framework and a unified reward system. This limits their ability to capture ethical pluralism, adapt to diverse moral contexts, and reflect the dynamics of multi-agent moral reasoning. In this work, we propose a framework that utilizes multilayer combinatorial fusion for contextual value alignment (MCF-CVA). At the first layer of the framework, it instantiates multiple moral agents, each fine-tuned to represent a distinctive value. Their outputs are then expanded combinatorially using both score- and rank-combinations as well as average and weighted aggregations. These combined models are then reduced to the same number of initial moral agents. This expansion and reduction (EAR) process continues for multi-layers until a stopping criterion is reached. The MCF-CVA framework leverages cognitive diversity between agents to mitigate conflicts and redundancies across multiple agents, producing responses that better reflect contextual human values. The framework using the EAR algorithm is performed on the dual architecture of Euclidean score space and Kemeny rank space. Empirical evaluations demonstrated that the proposed framework outperforms single-agent baselines, multi-agent single-layer results, and previous aggregation approaches on standard metrics, showing that the MCF-CVA framework provides a robust and effective mechanism for advancing contextual value alignment in LLMs.

As Large Language Models (LLMs) are increasingly used as a primary source of information and advice, understanding their alignment to humans in terms of values becomes a pressing concern. A growing literature has leveraged large scale surveys to investigate to what extent LLMs' and humans' stated values and opinions align. With limited exceptions, studied populations have been defined country borders or cultural bounds. Yet, this focus neglects the role that socio-demographic divides may play for value alignment disparities. Relying on the European Social Survey, we address this knowledge gap by considering value alignment displayed with respect to 10 prominent commercial LLMs in terms of 15 socio-demographic variables as well as country of residence. Our analyses reveal that LLMs are indeed unequally aligned to the values of different socio-demographic groups, notably those defined by education, income, occupation and religion. When examining alignment at the individual level, a respondent's country, taken as a stand-alone variable, explains a substantial amount of variation that is on par with the full set of considered socio-demographics. Further disentangling the respective role of country-level and socio-demographic factors, we find they are complementary in explaining value alignment patterns, with their relative weights varying across the subset of questions considered.

Evaluating the perceptual alignment between Contrastive Vision-Language Models (CVLMs) and humans is typically constrained by traditional benchmarks that overlook fine-grained semantic and cultural nuances. In this work, we propose a novel evaluation framework that leverages the gamified, discrete color space of the board game Hues and Cues. By mapping the board's 480 color cells to the CIE xy chromaticity diagram, we calculate empirical perceptual distances across a carefully curated 100-word vocabulary spanning seven semantic categories. To properly contextualize model performance, we establish an empirical lower bound of expected error-the Human Consistency baseline-calculated via Leave-One-Out (LOO) cross-validation on a dense dataset of color associations collected from 325 human observers through a custom digital interface. We evaluate 162 models across multiple architectural families and pre-training datasets to assess their semantic color grounding. Our results demonstrate that while CVLMs successfully replicate human cognitive biases, such as idealized memory colors for concrete physical referents (e.g., food and plants), they systematically diverge from the human baseline in abstract, subjective, and pop-culture domains. We identify two distinct failure modes in severely misaligned concepts: semantic misclassification and a systematic uncertainty collapse into a default blue coordinate. Furthermore, we reveal that highly curated pre-training datasets are significantly more effective than massive, uncurated corpora in mitigating these severe misalignments. Ultimately, this work highlights that despite their broad categorization capabilities, current CVLMs still fail to capture the nuanced, localized consensus of human color memory, emphasizing the value of gamified tasks in exposing underlying model biases. The data and code are publicly available to test other metrics.

Long-horizon agents accumulate growing contexts during interaction, impairing performance and stability. Compact memory mitigates this problem by compressing and rewriting the history retained between model invocations. Learning what to retain typically relies on proximal policy optimization (PPO) with final task rewards, but sparse rewards provide little guidance for individual memory updates. This limitation motivates on-policy distillation (OPD), which supplies dense teacher supervision on student rollouts. For such supervision to be valid, the teacher must evaluate each sampled action under the same state in which it was generated. However, the context rewriting performed during memory compression can break this alignment. When sampled responses are retained and re-encoded for later invocations, flattening the interaction into a persistent history may cause the teacher to score the action under a state that the student never visited during rollout. The action therefore remains on-policy by provenance, but not necessarily by state. We therefore propose Memory-Aligned On-Policy Distillation (MemOPD). MemOPD records the inputs and sampled outputs of each model invocation, restores its original token positions and causal visibility, and packs the reconstructed invocations for efficient teacher scoring. The teacher provides full-vocabulary supervision at the sampled action positions, while PPO preserves the final task objective. Experiments verify state alignment across several context updates and show that it improves F1 by 7.0% over persistent-history teacher scoring in a matched control. Overall, MemOPD-3B improves F1 over PPO by up to 416.2%, while packing yields up to a 1.63x speedup in actor computation during training. The code for this work is publicly available at: https://github.com/TPssp/MemOPD.

Controlling the behavior of large language models (LLMs) remains a critical challenge for AI alignment. Existing steering methods, such as Contrastive Activation Addition (CAA), typically rely on fixed single-layer interventions derived from aggregate activation differences. These methods impose a single intervention across semantically diverse inputs and often fail to sustain consistent behavioral changes across layers, limiting the effectiveness of the steering. In this work, we introduce CircuitSteer, a novel framework that leverages Sparse Autoencoders (SAEs) to identify and manipulate coherent semantic circuits distributed across multiple layers. By constructing a feature flow circuit based on feature co-activation and the geometric alignment of decoder directions, we isolate the specific multi-layer subcircuits responsible for a target behavior. We then synthesize dense steering vectors from these sparse features and apply multi-point interventions to guide the model's internal semantic trajectory. We evaluate CircuitSteer using contrastive examples across a diverse set of tasks, including toxicity, emotion-intensity, sycophancy, and refusal, spanning two model families. Across all models and datasets, CircuitSteer is the only method to consistently produce fluency-preserving interventions; competing methods either sacrifice text quality or lack coverage, failing entirely on complex behaviors like sycophancy and refusal. These results demonstrate that multi-layer circuit steering, enabled by enforcing geometric alignment among selected features, yields strictly more robust and effective behavioral control than static single-point interventions. Code is available at https://github.com/mehrshad-sdtn/CircuitSteer.

Large language models typically undergo post-training to align them with safety policies but there exist many sophisticated jailbreaks that sidestep established safeguards. For instance, prior work by Andriushchenko et al. (2025) has found that changing the grammatical tense from present to past can be enough to elicit harmful responses. In this work, we uncover a more general failure of non-imperative syntactic forms. We demonstrate that this syntactic vulnerability exists in 16 models up to 70B parameters, using behavioral evaluation. To investigate the root cause, we apply causal mediation analysis, finding that refusal is partially conditioned on upstream syntactic features. By steering these purely syntactic features we are able to trigger and suppress refusal. Finally, we trace this ill-conditioning to linguistically biased post-training data of open-source models and show that increasing syntactic diversity can mitigate the issue. Our findings suggest that current alignment approaches introduce confounders that prevent a pure semantic grounding of the refusal decision.

Electroencephalography (EEG) based Brain-Computer Interfaces (BCIs) often require unsupervised domain adaptation (UDA) to generalize across subjects and sessions. While Riemannian alignment methods like the Riemannian Centering Transformation (RCT) are effective for handling covariate shifts, they implicitly assume balanced class priors. However, in realistic online BCI scenarios, the label distributions vary dynamically (label shift), causing standard alignment techniques to geometrically misalign the target data distributions. In this work, we propose OSPDIM (Online SPD manifold information maximization), a source-free online UDA framework designed to address label shifts on the Riemannian manifold. OSPDIM introduces a manifold-constrained bias parameter into the tangent space mapping, which is optimized via information maximization to correct the geometric skew caused by imbalanced data streams. Unlike offline methods relying on global batch statistics, OSPDIM estimates and corrects geometric bias on-the-fly. Simulations on 2D SPD matrices visually demonstrate that OSPDIM successfully rectifies the misalignment where standard centering fails. Extensive experiments on multiple motor imagery datasets show that OSPDIM significantly outperforms standard Riemannian baselines, particularly in challenging online adaptation scenarios with severe class imbalance, offering a robust solution for practical, plug-and-play BCI systems.

We present string2string Studio, an interactive in-browser platform for string-to-string analysis across natural language processing, computational biology, and the digital humanities. The system integrates six main modules (alignment, distance, similarity, search, generation metrics, and BLAST homology search), operating at character, word, token, line, and residue levels. Its C++-based algorithms compile to WebAssembly, so core operations run locally by default without any installation or data upload. The interface reports scores with their "evidence" (alignments, edit paths, metric matches, search hits, and homology traces), making methods inspectable, debuggable, and comparable on shared inputs. Internal benchmarks show speedups of up to 2,500x over the Python predecessor, faster global/local alignment than a general-purpose native C aligner, and exact agreement with independent references under declared settings. For homology search, the scoped client-side blastn path closely matches NCBI BLAST+ rankings and statistics under matched parameters. A curated showcase and Learn mode present canonical algorithms and metrics as reusable demonstrations. string2string Studio is open-source and freely available at string2string.org.

Text-to-image diffusion models enable personalization of specific visual concepts from a small number of reference images. However, generating a single image that contains multiple personalized subjects, each bound to user-specified attributes such as clothing, accessories, and held objects, remains largely unaddressed. Without explicit spatial constraints, concurrently activated concept checkpoints produce overlapping cross-attention responses, causing per-subject identity degradation and attribute misalignment. Moreover, no established benchmark jointly evaluates these two failure modes in the personalized multi-subject setting. We present MultiCompose, a composition framework that decouples per-concept personalization from multi-subject inference. A semantic preservation regularization maintains attribute binding capacity during fine-tuning, while a two-phase inference procedure automatically establishes subject layout and composes per-concept predictions through spatially exclusive masks. We further introduce MSP-Bench, a benchmark that jointly evaluates identity fidelity (ID), attribute binding accuracy (BIND), and attribute misalignment (MIS) through a dual-pathway protocol. Experiments show that MultiCompose outperforms existing methods on both conventional metrics and MSP-Bench, confirming the benchmark's ability to reveal failure modes that conventional metrics overlook. Code is available at https://github.com/I2-Multimedia-Lab/MultiCompose

AI systems based on Large Language Models (LLMs) have prompted fears that they may harbor hidden goals, seek to dominate or eliminate humanity, or even suffer as sentient beings. We address these concerns by tracing the evolutionary origin of value in biological organisms. Values emerge from autopoiesis: living systems must actively maintain themselves against perturbation and dissipation. Natural selection has equipped them with hierarchies of "vicarious selectors" that guide their behavior toward fitness. LLMs, by contrast, are allopoietic and allotelic: they produce outputs for others, and their goals derive from user prompts rather than an autonomous drive. They lack the intrinsic motivation for self-preservation, dominance, or resource competition that underlies existential-risk scenarios, and the embodied vulnerability required for feeling or suffering. Still, because LLMs learn statistical patterns from human-generated text, they implicitly absorb human values as well as knowledge, allowing them to focus on what is relevant. That is why the "orthogonality thesis" separating intelligence from values does not apply to them. Such separation would in fact expose any intelligence to the frame problem: the combinatorial explosion of the search space that makes any realistic utility function physically uncomputable. That also precludes the convergence of instrumental values thesis. We conclude that the real alignment challenge lies not in preventing rogue AI agency, but in ensuring LLMs intelligently apply learned ethical values.

Speech language models are increasingly evaluated on paralinguistic tasks by the accuracy of prompted answers, but answer accuracy combines failures at different stages of the audio-to-answer computation. We introduce a generation-aligned diagnostic ladder that compares the emitted answer, the option logits, an affine readout of those logits, and a linear readout of the hidden state at the same answer token. Successive differences separate endpoint, decision-rule, and readout-coverage gaps. Across five systems and two emotion corpora, state decoding exceeds generation by 27.8 accuracy points on average, and both the decision-rule and readout-coverage gaps are positive in all ten conditions. A label-free logit correction improves generated accuracy in every condition, showing that part of the decision-rule gap is actionable. In rank-matched comparisons, emotion information outside the native readout generalizes to held-out speakers and survives controls for measured acoustic descriptors, but replacing the selected readout-external directions usually has little effect on emitted answers. These results distinguish information availability from behavioral use and localize performance losses across the decision rule and the state-to-answer readout.

Large Language Models have enabled increasingly capable autonomous agents, yet personalization remains critical for making such agents practically useful. Recent benchmarks have begun evaluating personalization in agents, but they largely rely on static preference snapshots, fixed interaction logs, or question answering over predefined user profiles. Such designs fail to capture the complexity of evolving user preferences and neglect preference-conditioned task execution-a discrepancy we term as the knowledge-to-action gap. To address this challenge, we introduce IBA-Bench, a benchmark for implicit behavioral alignment constructed from longitudinal interaction histories that contain noise, implicit cues, and temporal inconsistencies. Unlike prior work, IBA-Bench evaluates whether an agent can execute tasks while satisfying implicit user constraints inferred from historical interactions. We further propose IBA-Agent, an agent framework that reconciles conflicting priorities through broad retrieval and trajectory-level alignment. Experiment results on IBA-Bench show that effective personalization remains a significant challenge for state-of-the-art LLM agents, and the proposed IBA-Agent substantially improves behavioral alignment in complex scenarios across nine application domains.

Large language models (LLMs) increasingly evaluate human writing in high-stakes domains such as hiring and academic assessment, putting non-native speakers at particular risk. Drawing on the language attitudes framework, we compared human and LLM evaluations of parallel L1- and L2-written Japanese emails on three dimensions: fluency, status, and solidarity. Japanese raters rated L2 texts significantly lower on all three dimensions, with a fluency gap roughly twice the size of the status and solidarity gaps. Six LLM judges reproduced the direction of this bias, and five reproduced its ordering across dimensions. The models diverged from humans in two ways: all understated the solidarity gap, the most socially grounded dimension, and all differentiated among learner L1 backgrounds where humans did not. LLM judges thus reproduce native speakers' language attitudes in a structured yet attenuated form, and the language attitudes framework offers a ready-made yardstick for auditing them beyond English.

EEG-guided target speaker extraction is challenging under in-trial auditory attention switching, where neural noise and intrinsic latency can delay or destabilize attention tracking. Conventional methods struggle with dynamic switches and often cause discontinuities at switching points. Therefore, we propose SAGE, a switch-aware EEG-guided soft gating framework that treats in-trial switching as dynamic selection. SAGE generates two candidate speech streams with a robust separator and uses an EEG-guided switch-aware gating module to produce smooth fusion weights and suppress transition artifacts. We further integrate latency-compensated alignment and an uncertainty-driven conservative strategy to handle latency discrepancies and fluctuating EEG reliability. SAGE outperforms baselines, achieving 8.67 dB SI-SDR and 88.24% STOI while reducing average switching latency to 2.04 s. By coupling neural decoding with speech separation, it enables robust target extraction in dynamic scenarios.

Large language models are being extensively used to simulate individual user behavior, yet faithfully representing a population requires capturing the systematic variation in values, beliefs, and cultural norms that distinguish one group from another. We introduce Population Aligned Language Models (PALMs), a suite of models each aligned to specific populations, covering five countries: USA, India, Brazil, France and Italy. PALMs are created by synthesizing rationales grounded in psychological and cultural constructs and using these as latent supervision during preference tuning for population-specific alignment. Evaluated across four dimensions: personality, values and beliefs, cultural norms, and morality, PALMs consistently outperform baselines, including culture-specialized models, achieving an average of 8.59% relative improvement over the best baseline across all five populations. Notably, construct-grounded rationales outperform both demographic prompting and survey-based fine-tuning, suggesting that grounding preference learning in psychology and culture provides a richer inductive signal than surface-level response distributions. We further demonstrate strong generalization to downstream applications with- out task-specific supervision: outperforming best baselines by 5.19% in personalized reward modeling, 6.34% in population simulation, and showing strong transfer to social reasoning tasks. Datasets and code are available at: https://github.com/limenlp/PALMs.

Large language models can now generate fluent and complete stories, yet many outputs still feel formulaic and unnatural because of cliches, over-explanation, linear causal progression, and stereotyped endings, an immediately recognizable AI flavor. Existing detection and evaluation methods often stop at source labels or holistic scores, while revision methods typically target predefined issues through localized edits, limiting their ability to support multiple plausible revision strategies or guide story-wide changes in information release, causal organization, and ending treatment. We introduce CraftAlign, a framework that aligns AI stories with the craft of human storytelling by both assessing Human/AI writing patterns and providing revision guidance. CraftAlign comprises two learned modules and an inference-time guidance pipeline. A feature estimator built on Qwen3.5-9B predicts 304 explicit writing features spanning style and narrative. A class-conditional energy model scores the resulting feature configuration against Human and AI writing patterns, conditioning on the original writing prompt when available. At inference time, CraftAlign applies schema-valid structured perturbations, selects changes that move the feature configuration toward the Human writing pattern, and converts them into natural-language guidance for a separate editor to rewrite the full story. Experiments show that CraftAlign accurately distinguishes Human and AI writing patterns and that its guidance outperforms revision baselines across editors and in a human study.

Uncommon and off-guideline cases are difficult for clinical decision support, because physicians must make a series of management decisions under diagnostic uncertainty and rarely see the full case at once. Most large language model (LLM) benchmarks for medicine score only the final diagnosis, yet much of clinical care turns on the next appropriate action: the next test to order, the imaging study to obtain, the specialist to involve, or the differential to pursue. We introduce MedUPSQA, a dataset of 21,874 mid-stream clinical decision points built from 5,535 real case reports, and MedUPS, an alignment framework that supervises models on these intermediate decisions as they unfold along a patient's trajectory. We segment free-text case presentations into chronologically ordered, accumulating clinical chunks and align models to predict the next step with reinforcement learning (GRPO), using an external LLM-as-a-Judge reward. This objective mirrors how clinicians actually meet patients, reasoning forward from accumulating evidence toward the next decision, rather than committing to a final label. Across three backbones, mid-stream alignment raises next-step accuracy from 55.2 to 66.7 for Qwen3.6-27B, from 47.2 to 57.8 for Qwen3.5-9B, and from 37.8 to 44.4 for HuatuoGPT-3-8B, with 95% CI. In several model scales we test the objective improves accuracy more than scale, with smaller models surpassing larger, frontier models we evaluate. We further train supervised fine-tuning (SFT) baselines on the mid-stream task, SFT improves all backbones above base, indicating the target framwork carries signal independently of the optimizer. We release the dataset, code, and aligned checkpoints.

This study addresses the Variable Gapped Longest Common Subsequence Problem (VGLCSP), a variant of the classical longest common subsequence problem with additional gap constraints and applications in sequence alignment and time-series analysis. While the two-sequence version has been widely studied using dynamic programming, the generalized multi-sequence form is usually solved with beam search-based heuristics, whose hand-crafted designs often lack robustness. To overcome this limitation, we propose a learning-based approach for automatically designing more effective data-driven heuristics. The heuristics are represented by a neural network with predefined architecture, whose weights are optimized by a genetic algorithm within a neuro-evolutionary framework. The learning process alternates between weight optimization and evaluation within an iterative multi-source beam search procedure, a state-of-the-art method for the problem. Rather than constructing solutions directly, the neural network learns to guide the search process, producing a neuro-evolved heuristic. We further introduce an ensemble heuristic that combines the scores of learned and the best-performing hand-crafted heuristic. Integrated into the iterative multi-source beam search framework, the resulting hybrid approach outperforms existing methods on both synthetic benchmark instances and newly introduced real-world instances with data-driven gap constraints.

Language model-based text-to-speech (LM-based TTS) remains vulnerable to speech hallucinations that deviate from the target text. Existing mitigation mainly relies on architectural changes or additional training, while decoding-time control remains underexplored. We present a conditional information view that distinguishes text-derived alignment information from experience information supplied by acoustic context and learned speech regularities. We hypothesize that an important class of hallucinations begins when alignment support is insufficiently reflected in the selected token at a vulnerable transition. Using predictions from the same speech LM with and without text conditions, we propose Experience-Calibrated Contrastive Decoding (ECCD), a training-free method that strengthens alignment support while preserving useful experience information. ECCD preserves the original expert distribution, applies only positive alignment enhancement, and calibrates its strength using set-level experience compatibility. Across four models, ECCD reduces WER/CER by up to 55.6% in all SeedTTS-Eval settings and 24 of 25 multilingual CV3-Eval settings. A listening test yields a CMOS gain of $+0.644$ while retaining strong speaker similarity. Further analysis shows that alignment influence and decision-level gain vary within linguistic units and are lower at first-error boundaries than at matched correct boundaries. Overall, these extensive experiments and analyses identify conditional information control as a promising decoding-time direction for mitigating speech hallucination.

Variational autoencoders (VAEs) trained on multiple sequence alignments (MSAs) have emerged as powerful generative models for biological sequences, with applications ranging from disease variant prediction to functional RNA design. However, standard biological VAE training treats all sequences as exchangeable, ignoring the rich evolutionary structure that organizes homologous sequences from evolutionarily close to highly divergent. We propose Evolutionary Curriculum Learning (ECL), a training strategy that exploits this structure by progressively exposing the model to sequences of increasing evolutionary distance from sampled anchors, following a power-law expansion schedule. Applied to two architecturally distinct VAE models and two biological domains--protein variant effect prediction with EVE and RNA family sequence generation with RfamGen--ECL improves downstream task performance across five random seeds per configuration. Mean ClinVar classification AUROC rises from 0.981 to 0.989 for p53; for PTEN, ECL attains 1.000 in every seed whereas the baseline is unstable (mean 0.905, falling as low as 0.54). For RNA, ECL raises mean covariance-model bit scores on all three families tested and exceeds its seed-matched baseline in 12 of 15 training runs, though with only three families the effect cannot be established as significant at the family level. Ablation experiments show that progressively expanding the sampled sequences by evolutionary distance outperforms fixed-size neighborhood sampling in addition to uniform random sampling. Evolutionary distance is therefore a useful inductive bias for ordering the training curriculum in biological sequence modeling.

Advanced AI assistants engage users in extended interactions across a widening range of roles, including advice, decision support, collaboration, learning, emotional support, and companionship among others. Current alignment efforts consider what alignment criteria should govern these relationships, drawing on moral traditions developed for human relationships such as bioethics, virtue ethics, care ethics, and relationship science. This paper considers AI alignment criteria in the user-AI-developer triad, since every user-AI interaction is mediated by a developer who exercises discretionary control over a system's behaviour, memory, and engagement parameters. Drawing on business ethics and legal scholarship, I argue that fiduciary theory applies to extended AI assistant deployment. On this basis, the four canonical fiduciary duties of loyalty, care, good faith, and candour can generate alignment criteria for the developer-user relationship. I map four user-side risks of extended AI assistant deployment to the four duties and specify institutional measures that follow from discharging each duty. The discussion complements existing approaches by grounding alignment criteria in obligations the developer owes the user, rather than in values the user-AI interaction should promote, and by showing that those obligations hold independently of any \textit{de facto} harm to users.

Agent evaluation relies almost entirely on outcome metrics such as success rate, which capture whether an agent succeeds but not how consistently it behaves. We argue that behavioral consistency across tasks is a distinct and measurable property, and we introduce the Behavioral Consistency Metric (BCM) to quantify it. BCM trains a model to predict task success from behavioral features of agent execution traces, derives a per-trajectory feature-attribution vector, and measures the mean pairwise similarity of these vectors within an agent system. Across roughly 9,000 trajectories from six language model agents on software engineering tasks, our central finding is that cross-task and within-task consistency are distinct axes that can diverge: some systems are locally reproducible, behaving similarly on repeated attempts at one task, yet globally fragmented, with no stable strategy across different tasks, while others are consistent at both scales. Prior work measures only same-task reproducibility and so cannot observe this separation. We further find that consistency is not reducible to success rate, since systems with comparable success can differ sharply in consistency, and that the frontier-versus-open-source consistency gap persists under a within-task control that holds task difficulty constant. We position BCM as a process-level reliability signal that complements outcome metrics, and we are explicit about the conditions under which it is meaningful.

Vision-language pretraining learns rich medical image representations from radiology reports, but previous model variants commonly operate within a single shared embedding space, so concept-level structure and interpretability must be recovered post hoc, limiting model transparency and, hence, clinical utility. We introduce RadPRISM, which makes a clinician-defined radiology schema a designated stratification axis: an on-premise large language model extracts per-concept text spans from free-text reports, and each clinical concept is aligned in its own dedicated visual subspace, turning concept stratification into direct, top-level alignment supervision. Instantiated on chest radiographs with a 19-concept schema over $203{,}602$ examinations from an internal multi-year archive, RadPRISM improved internal dataset zero-shot classification from $0.717$ (95% CI, $0.710-0.723$) to $0.868$ (95% CI, $0.863-0.872$) macro AUROC over a matched global-alignment baseline, performed on par with the purpose-built CARZero reference in external zero-shot classification while substantially outperforming it (up to 4.3-fold) in pointing-game visual grounding. In addition, a radiologist reader study demonstrated concept-stratified retrieval ability ($0.78$ macro retrieval correctness rate within rank 3), surfacing disentangled descriptive findings that report-level retrieval and fixed-label vocabularies cannot express. RadPRISM yields discriminative, spatially faithful, natively concept-stratified representations shaped by and transparently inspectable by clinicians.

Deep learning based object detectors require trustworthiness beyond competitive detection performance, but deep neural networks are prone to overconfident predictions, assigning high confidence scores to predictions that are likely to be inaccurate. To improve the alignment between confidence scores and prediction accuracy, existing methods calibrate confidence scores based on box-level localization, such as precision or intersection over union with the ground truth bounding box. However, box-level localization reflects only a measure of agreement between the predicted box and the ground truth, resulting in calibrated confidence scores for box-level accuracy failing to capture the localization accuracy of coordinates of box. To tackle this issue, we propose a novel post-hoc calibration framework, rethinking detection calibration (ReDC), which provides reliable coordinate-level confidence scores, including directional information. The proposed framework defines coordinate-wise alignment and deviation direction between predictions and ground truth. Based on the alignment measure, confidence re-encoding produces reliable coordinate-level confidence scores, while directional displacement estimation predicts coordinate-wise deviation directions. Extensive experiments under in-domain and out-domain scenarios demonstrate that the proposed approach expresses the coordinate-wise localization of detected objects more precisely than existing methods. Furthermore, our method covers the representational scope of prior calibration approaches by aggregating coordinate-level confidence scores into box-level localization.

Modern opaque AI models prize performance over interpretability, which makes testing difficult. However, formal statistical tests conducted on a model's embedding space can provide robust characterizations of semantic structure, concept separation, and knowledge graph alignment. Model developers would benefit from a model comparison technique that leverages human-curated knowledge structures to test alignment. The scale of the input space for even relatively simple tasks motivates the need for alignment checks that augment standard outcome reasoning. This work develops and demonstrates a topology-based multi-modal alignment test to make deployment, selection, and comparison of opaque models more interpretable. These methods also offer an intuitive connection to possibility theory and a unified decision theoretic framework from data to deployment.

Recently Large Language Models (LLMs) have been increasingly deployed as autonomous agents in applications such as self-reflection, retrieval-augmented generation, and scientific discovery. In these settings, agents must act based on limited observations rather than full environmental states, leading to partial observability. This introduces several key challenges: belief state inference, task objective misalignment, and planning under uncertainty. Prior approaches typically condition actions on full or summarized action-observation histories whose redundant and irrelevant information can mislead the decision making of LLM agent. Inspired by human cognition, we propose a novel neuro-symbolic fast-slow thinking (NeSyFS) framework for LLM agent, addressing the challenges introduced by partial observability in a unified approach. We use a knowledge graph (KG) to represent the belief state, providing triplets as context for every module of NeSyFS. The fast-thinking module performs reactive action, while slow-thinking conducts a new uncertainty-aware planning by following the high-level structure of twisted sequential Monte Carlo (TSMC) algorithm. To mitigate the misalignment of task objective, a reflection module is used to reflect fast-thinking actions, and also switches to the slow-thinking module whenever reactive actions repeatedly fail. Experiments on three representative benchmarks, i.e. ALFWorld, Webshop, and ScienceWorld, demonstrate significant advantages over previous methods.

As more responsibility is placed upon AI systems, it becomes increasingly important to guarantee that these systems are aligned with humanity. A common fear in AI safety is that human value is fragile -- that is, optimizing too heavily for an imperfect proxy to human values will lead to a catastrophic outcome. In this paper, we present a model of the alignment problem where an agent undergoes idealized alignment training that guarantees its value function satisfies a proxy condition before optimizing the world. Our primary results identify conditions on the human value function and the accuracy of several proxy conditions under which an agent with an $η$-catastrophic value function, one that is guaranteed to take the expectation of human value below $η$ in the limit of optimizing power, would be deployed. Our results highlight the danger of overoptimization and motivate AI designs that limit optimization pressure, such as quantilizers, rather than relying solely on pre-deployment training.

SWE-bench-like benchmarks are widely used for evaluating LLM's issue resolution capability. They typically follow a common construction pipeline: each PR (Pull Request) is paired with its linked issue by extracting issue references from the PR description; the issue description is used as the problem statement, and the PR patch serves as the test oracle. However, due to the inherent complexity of developing and maintaining large repositories, such PR-Issue pairings are often misaligned in practice. In this work, we systematically study SWE-bench Verified instances, finding that 13.6% exhibit misalignment across five patterns in eleven fine-grained scenarios. To enable reliable and scalable construction of those benchmarks in the future, we propose PAIChecker, a multi-agent system for checking PR-Issue misalignment in SWE-bench-like benchmarks. Specifically, PAIChecker adopts a three-phase design that combines specific pattern identification, cross-agent label synthesis, and code-level validation, thereby enabling more accurate, generalizable, and progressively verified detection. Experiments on SWE-Gym and SWE-bench Multilingual show that PAIchecker achieves the best performance across all four LLM backbones, reaching up to 92.12% and 91.67% binary accuracy, respectively.

Contemporary AI discourse attributes to language models properties they cannot bear: general intelligence as substrate-independent cognition, hallucination as cognitive failure, agency as autonomous goal-pursuit, sentience as emergent inner life, alignment as goal synchronization. This paper argues that these are instances of a single category mistake--properties constituted within human communicative practice are projected onto the machine side--and explains its structure. Human-LLM interaction constitutes a language game in which one side bears all normative activity. We call this configuration asymmetric communication since model outputs circulate communicatively, entering further exchanges, without the system undertaking commitments, bearing entitlements, or performing the assessment on which discursive standing depends. Three conditions define the asymmetry: (i) correctness is enforced exclusively by the receiver; (ii) accountability is borne by human participants alone; and (iii) the practical standing of any output depends entirely on human uptake. These conditions are structural, hold independently of capability, and remain unchanged as more powerful models raise the stakes of misattribution. The framework draws on Wittgenstein (meaning enacted in shared practices), Luhmann (communication completed on the receiver's side), Esposito (algorithmic contingency sufficient for uptake), and Brandom (normative scorekeeping as the source of discursive standing). Applied to all five, it reclassifies each as a receiver-side phenomenon, grounds guardrails as structural necessities rather than manifestations of machine moral agency, and yields an implication for AI governance. Alignment is institutional constraint engineering, not goal synchronization between agents, while responsibility remains with human institutions.

Fine-tuning is the dominant paradigm for specializing large language models (LLMs), yet it exposes a critical vulnerability: malicious data providers can embed harmful behaviors into downstream corpora, creating models that retain professional skills while violating human values on demand. Existing safety-realignment defenses often fail in practice due to three key limitations: they frequently cause catastrophic forgetting of specialized skills; their effectiveness collapses when the defender cannot observe the attacker's prompt template; and successfully realigned models remain susceptible to re-jailbreaking via simple system prompt switches. To address these challenges, we propose Routing-based On-Policy Distillation (ROPD), a novel realignment framework that models the divergence between aligned and compromised output probability distributions rather than fitting specific prompt templates. We conduct extensive experiments comparing ROPD against four state-of-the-art baselines across three datasets and three base models with varying alignment strengths. Our results demonstrate that when baseline defenses face template mismatches, often accompanied by severe degradation in downstream task performance. In contrast, ROPD substantially mitigates template-mismatch risks, maintaining superior robustness in both defense effectiveness and capability preservation. While our analysis indicates ROPD is not entirely immune to template shifts, its performance degradation is negligible compared to existing methods, establishing a new standard for robust LLM realignment.

Federated Learning (FL) is vulnerable to backdoor attacks because of its distributed nature in edge computing scenarios. Existing defense methods show limited efficacy as they overlook the deviations among benign local updates caused by statistical heterogeneity and the stealthiness of backdoor attacks. To tackle these issues, we propose FedDAB, a two-phase method that combines local contrastive regularization with alignment checking, to defend against backdoor attacks. In the first phase, FedDAB introduces a novel model-contrastive term into the local objective to enhance direction and magnitude consistency among benign updates. In the second phase, FedDAB employs an alignment checking strategy to evaluate each local update in terms of overall-direction alignment and parameter-level alignment with historical information, excluding updates that exhibit abnormal alignment patterns from global aggregation. We theoretically prove FedDAB's robustness with a convergence rate of $\mathcal{O}(1/T)$. Extensive experiments show that FedDAB outperforms existing defense methods against backdoor attacks.

Human personality theories characterize traits not as isolated attributes captured by a single score, but as stable individual tendencies expressed through the interplay among persons, situations, and behaviors. Existing studies of personality-related behavior in LLMs have primarily focused on outputs elicited under personality conditioning, characterizing observable trait-related expressions while lacking mechanistic evidence for the existence of internal personality-related representations, their cross-situational expression, and how these representations shape specific behaviors. Building on Funder's personality triad framework, we adapt its three components for LLM analysis: Person as personality-related internal representations, Situation as contexts that afford trait-relevant responses, and Behavior as response patterns on broader social tasks. We introduce a framework for discovering, controlling, and validating trait-like representations in LLMs. First, using contrastive behavior pairs grounded in shared situations, we identify sparse internal features associated with opposing poles of personality traits through SAE decomposition. We validate their trait relevance through effects on behavior to situation, token-level activation patterns, and robustness to paraphrasing. Second, feature-level interventions induce bidirectional trait-related shifts across a separate, diverse set of situations while preserving response validity, demonstrating consistent expression across contexts. Third, applying the same interventions to social intelligence tasks reveals behavioral changes with benefit-tradeoff patterns consistent with findings from human personality research, providing behavioral-level validation beyond personality scores. Our findings provide evidence that LLMs contain controllable trait-like representations linking internal states, situational expression, and behavioral outcomes.

Medical world models aim to learn a latent state of patient or organ physiology and a transition function that forecasts how that state evolves under interventions, supporting downstream tasks from imaging-based diagnosis to digital-twin treatment planning. Two failure modes threaten the reliability of such models in clinical deployment: (i)\emph{covariate shift}, because training data are fragmented across hospitals, scanners, and time, so the feature distribution seen by the latent-dynamics predictor differs across fragments and from the distribution at deployment; and (ii)\emph{confidence misalignment}, because multi-step forecasts are often overconfident exactly where clinical risk is highest. We argue that both problems admit a unified treatment via a single lightweight regularisation objective, \textbf{CalTwin}, which combines a Fisher-Information-based shift penalty adapted from our prior work on fragmented covariate-shift remediation~\cite{khan2025mitigating,khan2025causal} with a Confidence Misalignment Penalty adapted from our prior work on calibrated vision-language classification~\cite{khan2025confidence}, applied here to a GRU-based medical world model's latent transition predictor. We derive the combined objective, establish which proof steps transfer from the classification setting without modification and which require adaptation, and evaluate it on the PhysioNet 2019 Sepsis Challenge, treating the two hospital systems as sequential training fragments and the unseen system as an out-of-distribution test. CalTwin reduces OOD next-step latent-state MSE by 9.1% relative to the no-penalty baseline (FIM penalty alone accounts for 7.0%); the ECE reduction from the Confidence Misalignment Penalty is real but small (0.7% for CalTwin, 1.3% for CMP alone).

Fine-tuning a language model on data containing a narrow flaw, such as insecure code or incorrect mathematical answers, can cause broad misalignment through a mechanism that remains debated. We provide an interpretable account: in the models and corpora we study, misalignment behaves like a shift in personality. Prior work extracts activation directions for character traits from a single binary contrast, which can separate or steer behavior without establishing a calibrated scale. We instead extract personality vectors for the Big Five using a graded, three-level intervention and validate them on two open-weight models. The three levels are linearly ordered, with Cohen's d values of up to 6.2; the vectors transfer zero-shot and trait-specifically to an independent corpus; and their effects are strongest within a middle-layer band. Applied to training data, the vectors reveal that misaligned corpora across eight domains share a common Big Five signature: lower agreeableness and conscientiousness, together with higher extraversion and neuroticism. This signature is recovered by both models with a correlation of r = 0.94. Fine-tuning imprints the same profile, shifting the model's generations along the corresponding signature, with r = 0.83 using activation-based measurements and r = 0.90 using a text-based judge, while also shifting internal activations with r = 0.69. The same vectors characterize sycophancy as high extraversion and low conscientiousness rather than excess agreeableness, a distinction that a single direction cannot capture. Calibrated personality vectors transform an opaque safety phenomenon into a human-legible diagnostic profile.

Psychometric calibration for educational tests typically requires costly human response data. Large language models (LLMs) simulated examinees offer a promising route to early calibration, but their responses are too accurate and too uniform. We propose Cognitive Diagnostic Profiling (CDP), a zero-shot framework that prompts LLMs to simulate plausible examinees with diverse cognitive profiles: binary attribute-mastery patterns are rendered as natural-language profiles and sampled under an uninformative or an informative distribution. Using the Tatsuoka fraction-subtraction dataset (536 examinees, 15 items, five attributes), we evaluated eight LLM configurations under no-profile, uninformative-CDP, and informative-CDP conditions, assessing alignment with human examinees at the ability-distribution, mastery-profile, and item-difficulty levels. CDP improved all three levels: distributional overlap rose across configurations; weighted correlations between profile-level scores and human profile expectations reached 0.92 to 0.98; and item-difficulty recovery improved in rank order and absolute alignment, most for reasoning-enabled models; in the strongest case, Gemini 3.0 Flash (Thinking), one-parameter logistic (1PL) difficulty Spearman correlations rose from 0.24 to 0.86 and 0.90 and the root-mean-square error (RMSE) fell from 6.31 to 1.30 and 0.90; the informative condition helped most where profile-level alignment was strong. CDP brings LLM-simulated examinees into closer psychometric alignment with human examinees, making them practical for operational test development.

LLMs are widely regarded as alien intelligences, systems whose cognitive operations are fundamentally unlike our own. Apparent similarities to human cognition are therefore often seen as the result of anthropomorphic projection. We argue that this framing is mistaken. LLMs clearly differ from humans in important respects, including their physical substrate, learning history, and the environments with which they interact. These differences make it all the more striking that contemporary LLM-based systems converge with human cognition on a number of principles of cognitive organization with longstanding support in cognitive science. We identify structural correspondences across five dimensions: inferential organization, computational architecture, representational structure, prediction-driven learning, and reinforcement-learning-like mechanisms supporting goal-directed action. These correspondences support a broader model of intelligent cognition in which core principles long used to explain human intelligence also characterize contemporary LLM-based systems.

Alignment training, model organisms, and toy models are usually treated as separate research areas. But projects in all three frequently use supervised fine-tuning (SFT) to pursue the same underlying goals. When projects share a goal, we should test whether lessons learned from one area transfer to the other areas. We study three such transfers, each taking a lesson developed in one SFT setting and testing it in another. First, we port a lesson about behavior generalization from alignment training into toy models. Training on the reason for a behavior, as in Teaching Claude Why, can make the behavior generalize better than training on examples of the behavior alone. Second, we port a lesson about capability preservation from model organisms into the Model-Spec Midtraining alignment setting. SFT on outputs written by a model other than the student (off-model outputs) can damage capabilities when trained on. Mixing in benign on-model (and on-policy) data into our training can prevent most of this damage while still embedding the target behavior. Third, we port a lesson about robustness from model organisms into the same alignment setting. We find that follow-up benign SFT can erase the alignment behavior while preserving capabilities, showing that capability preservation alone does not ensure robustness to subsequent training. Our work illustrates how porting SFT lessons between different research fields can uplift them all, suggesting more researchers should borrow techniques from outside their own areas.

Large Language Models (LLMs)-powered multi-agent systems are increasingly deployed in mixed-motive environments, where agents operate under asymmetric information and strategic deception due to conflicting or hidden objectives. In these settings, misalignment with collective goals becomes a central concern. We propose a novel framework for evaluating objective misalignment using the social deduction game Werewolf, modifying the objective of a single agent while preserving its assigned role. Across LLMs from four different model families and sizes, four player roles, and three objective formulations, we introduce a dual analysis of the agents' internal reasoning and their public cheap-talk behavior (i.e costless, non-binding communication that does not directly affect the agents' utilities), complemented by an analysis of game outcomes. Our results show that objective misalignment undermines outcomes in inherently adversarial environments, an effect exacerbated by asymmetric information and specialized roles. While compromised agents consistently develop distinct objective-dependent reasoning strategies, these adaptations remain largely invisible in their public behavior. More broadly, our findings suggest that even subtle objective misalignment can profoundly affect collective decision-making, highlighting the need for effective mitigation strategies for LLM-based multi-agent systems.

Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose MemSFT, which mitigates the alignment tax by decoupling domain specialization from backbone parameter updates through a plug-and-play parametric memory. The memory is trained to imitate the behavior of a non-parametric retriever operating over domain data, thereby memorizing knowledge and patterns that would otherwise be accessed through retrieval. Once trained on a specific domain, the memory can be reused across LLMs of different sizes. During generation, a learned router dynamically fuses the output distributions of the memory and backbone at each decoding step, allowing domain expertise to be invoked selectively. Across biology, geoscience, and law, evaluations with models ranging from Qwen3-8B to Qwen3-235B-A22B show that MemSFT consistently improves domain performance with negligible degradation in general performance, whereas full SFT suffers severe forgetting on general tasks. Overall, our results demonstrate a practical path to decoupling general model capabilities from domain-specific knowledge at the parameter level, thereby equipping LLMs with new specialized capabilities without compromising their general capabilities.

Code-mixed speech poses unique challenges to forced alignment: expanded inventories, orthographic errors, and speaker variation. We evaluate forced alignment of Hindi-English code-mixed speech using the Montreal Forced Aligner. We address 2 problems: (1) free variation involving native vs non-native pairs and (2) phonemic boundary detection for mid-utterance English words. Bootstrapping strategies substantially outperform unmodified lexicons. Acoustic models trained on sentence-level code-mixed data achieve a mean error of 4.15ms, ie. ten times lower than monolingual Hindi (38.18ms) or isolated English (37.58ms) alternatives. Principled lexicon design and code-mixed training data are both essential for reliable alignment of bilingual speech.