Paved with True Intents: Intent-Aware Training Improves LLM Safety Classification Across Training Regimes

We argue that safety classifiers should model user intent as an explicit signal between the prompt and the final label. To study this, we introduce AIMS, a human-annotated dataset of 1,724 difficult safety prompts, each paired with an intent description and harm label. We use AIMS to evaluate intent-aware training across supervised fine-tuning, preference learning, reasoning distillation, and reinforcement learning. Despite its size, AIMS enables competitive safety classifiers across training regimes: DPO from model-generated intent errors improves over SFT, and intent-conditioned distillation outperforms reasoning-only distillation in most teacher-student pairs. Most notably, directly rewarding intent faithfulness with GRPO yields the strongest average performance across five external safety benchmarks, while our intent-aware models form the inference latency-F1 Pareto frontier. These results show that faithful intent modeling is a compact, high-quality supervision signal for more robust safety classifiers.

Debiasing Without Protected Attributes: Latent Concept Erasure from Textual Profiles

Most fairness research in NLP assumes direct access to protected attributes such as gender, race, or nationality. In practice, however, such information is often unavailable due to privacy constraints, missing metadata, or legal restrictions, even though models may infer it from indirect textual cues. This raises a key question: can debiasing succeed without direct access to sensitive attributes? We propose H-SAL, which performs post-hoc concept and attribute erasure using self-description text as an implicit debiasing signal. To support this setting, we introduce a multi-domain Stack Exchange-based fairness benchmark for helpfulness prediction that includes both explicit and implicit signals, enabling comparison between standard debiasing with protected labels and debiasing without access to sensitive information. Across encoder and decoder-only language models, we find that implicit self-description often matches or outperforms explicit-label-based debiasing. Our results broaden representation-level fairness research and provide a new benchmark for studying debiasing under realistic data constraints.

When Knowledge Is Not Free: Cost-Aware Evidence Selection in Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) typically assumes that external knowledge is free, but many high-quality sources are paywalled, licensed, restricted, or otherwise costly to access. We introduce cost-aware RAG, a setting where retrieved evidence is assigned access-cost tiers and systems must answer under an explicit evidence-access budget. We instantiate this setting by augmenting MS MARCO v2.1 with access-friction tiers and evaluate budgeted evidence selection across general-domain and domain-specific QA benchmarks. Our results show that static selection is brittle: no fixed selector uniformly dominates, and larger budgets do not reliably improve answer quality, even when costly evidence is domain-matched. We then study agentic cost-aware RAG, where an LLM decides when to retrieve, which tier to access, and when to stop. Agents show strong promise as adaptive evidence-acquisition controllers, but their behavior remains highly model- and task-dependent. These findings suggest that cost-aware evidence acquisition is a central challenge for the next generation of RAG systems. All code and data are available at https://github.com/Mignonmy/Cost-Aware.

More Test-Time Compute Can Hurt: Overestimation Bias in LLM Beam Search

Wider beam search should improve LLM reasoning, but when should you stop widening? Prior work on beam width selection has focused on inference efficiency \citep{qin2025dsbd, freitag2017beam}, without analyzing whether wider search can \emph{hurt} output quality. We present an analysis, grounded in Extreme Value Theory, that answers this question. Beam selection over noisy scorer outputs introduces a systematic overestimation bias that grows with the candidate pool size, and we derive a maximum useful beam width $\hat{k}$ beyond which search degrades performance. This critical width depends on the signal-to-noise ratio of the scorer: $\hat{k}$ grows exponentially with $(Δ/σ)^2$, where $Δ> 0$ is the quality advantage of correct paths over incorrect ones and $σ$ is the scorer noise. We validate this theory by comparing perplexity-guided and PRM-guided beam search across three 7B-parameter models and ten domains on MR-BEN (5,975 questions). Perplexity scoring, with its high noise, yields $\hat{k} = 1$: search provides no benefit at any width tested. PRM scoring, with lower noise, yields $\hat{k} \geq 4$, with gains of up to 8.9 percentage points. The same model, the same algorithm, but different scorers place $\hat{k}$ at opposite ends of the beam width range. Our analysis identifies the scorer's signal-to-noise ratio as the key quantity governing beam width selection, and we propose diagnostic indicators for choosing the beam width in practice.

Self-Improving World Modelling with Latent Actions

Internal modelling of the world -- predicting transitions between previous states $X$ and next states $Y$ under actions $Z$ -- is essential to reasoning and planning for LLMs and VLMs. Learning such models typically requires costly action-labelled trajectories. We propose SWIRL, a self-improvement framework that learns from state-only sequences by treating actions as a latent variable and alternating between Forward World Modelling (FWM) $P_θ(Y|X,Z)$ and an Inverse Dynamics Modelling (IDM) $Q_φ(Z|X,Y)$. SWIRL iterates two phases: (1) Variational Information Maximisation, which updates the FWM to generate next states that maximise conditional mutual information with latent actions given prior states, encouraging identifiable consistency; and (2) ELBO Maximisation, which updates the IDM to explain observed transitions, effectively performing coordinate ascent. Both models are trained with reinforcement learning (specifically, GRPO) with the opposite frozen model's log-probability as a reward signal. We provide theoretical learnability guarantees for both updates, and evaluate SWIRL on LLMs and VLMs across multiple environments: single-turn and multi-turn open-world visual dynamics and synthetic textual environments for physics, web, and tool calling. SWIRL achieves gains of 16% on AURORABench, 28% on ByteMorph, 16% on WorldPredictionBench, and 14% on StableToolBench.

From Competition to Collaboration: Designing Sustainable Mechanisms Between LLMs and Online Forums

While Generative AI (GenAI) systems draw users away from (Q&A) forums, they also depend on the very data those forums produce to improve their performance. Addressing this paradox, we propose a framework of sequential interaction, in which a GenAI system proposes questions to a forum that can publish some of them. Our framework captures several intricacies of such a collaboration, including non-monetary exchanges, asymmetric information, and incentive misalignment. We bring the framework to life through comprehensive, data-driven simulations using real Stack Exchange data and commonly used LLMs. We demonstrate the incentive misalignment empirically, yet show that players can achieve roughly half of the utility in an ideal full-information scenario. Our results highlight the potential for sustainable collaboration that preserves effective knowledge sharing between AI systems and human knowledge platforms.

Lost in Space? Vision-Language Models Struggle with Relative Camera Pose Estimation

Vision-Language Models (VLMs) perform well in 2D perception and semantic reasoning compared to their limited understanding of 3D spatial structure. We investigate this gap using relative camera pose estimation (RCPE), a fundamental vision task that requires inferring relative camera translation and rotation from a pair of images. We introduce VRRPI-Bench, a benchmark derived from unlabeled egocentric videos with verbalized annotations of relative camera motion, reflecting realistic scenarios with simultaneous translation and rotation around a shared object. We further propose VRRPI-Diag, a diagnostic benchmark that isolates individual motion degrees of freedom. Despite the simplicity of RCPE, most VLMs fail to generalize beyond shallow 2D heuristics, particularly for depth changes and roll transformations along the optical axis. Even state-of-the-art models such as GPT-5 ($0.64$) fall short of classic geometric baselines ($0.97$) and human performance ($0.92$). Moreover, VLMs exhibit difficulty in multi-image reasoning, with inconsistent performance (best $59.7\%$) when integrating spatial cues across frames. Our findings reveal limitations in grounding VLMs in 3D and multi-view spatial reasoning.

Iterative Multilingual Spectral Attribute Erasure

Multilingual representations embed words with similar meanings to share a common semantic space across languages, creating opportunities to transfer debiasing effects between languages. However, existing methods for debiasing are unable to exploit this opportunity because they operate on individual languages. We present Iterative Multilingual Spectral Attribute Erasure (IMSAE), which identifies and mitigates joint bias subspaces across multiple languages through iterative SVD-based truncation. Evaluating IMSAE across eight languages and five demographic dimensions, we demonstrate its effectiveness in both standard and zero-shot settings, where target language data is unavailable, but linguistically similar languages can be used for debiasing. Our comprehensive experiments across diverse language models (BERT, LLaMA, Mistral) show that IMSAE outperforms traditional monolingual and cross-lingual approaches while maintaining model utility.

Towards Large Language Models with Self-Consistent Natural Language Explanations

Large language models (LLMs) seem to offer an easy path to interpretability: just ask them to explain their decisions. Yet, studies show that these post-hoc explanations often misrepresent the true decision process, as revealed by mismatches in feature importance. Despite growing evidence of this inconsistency, no systematic solutions have emerged, partly due to the high cost of estimating feature importance, which limits evaluations to small datasets. To address this, we introduce the Post-hoc Self-Consistency Bank (PSCB) - a large-scale benchmark of decisions spanning diverse tasks and models, each paired with LLM-generated explanations and corresponding feature importance scores. Analysis of PSCB reveals that self-consistency scores barely differ between correct and incorrect predictions. We also show that the standard metric fails to meaningfully distinguish between explanations. To overcome this limitation, we propose an alternative metric that more effectively captures variation in explanation quality. We use it to fine-tune LLMs via Direct Preference Optimization (DPO), leading to significantly better alignment between explanations and decision-relevant features, even under domain shift. Our findings point to a scalable path toward more trustworthy, self-consistent LLMs.

Bootstrapping World Models from Dynamics Models in Multimodal Foundation Models

To what extent do vision-and-language foundation models possess a realistic world model (observation $\times$ action $\rightarrow$ observation) and a dynamics model (observation $\times$ observation $\rightarrow$ action), when actions are expressed through language? While open-source foundation models struggle with both, we find that fine-tuning them to acquire a dynamics model through supervision is significantly easier than acquiring a world model. In turn, dynamics models can be used to bootstrap world models through two main strategies: 1) weakly supervised learning from synthetic data and 2) inference time verification. Firstly, the dynamics model can annotate actions for unlabelled pairs of video frame observations to expand the training data. We further propose a new objective, where image tokens in observation pairs are weighted by their importance, as predicted by a recognition model. Secondly, the dynamics models can assign rewards to multiple samples of the world model to score them, effectively guiding search at inference time. We evaluate the world models resulting from both strategies through the task of action-centric image editing on Aurora-Bench. Our best model achieves a performance competitive with state-of-the-art image editing models, improving on them by a margin of $15\%$ on real-world subsets according to GPT4o-as-judge, and achieving the best average human evaluation across all subsets of Aurora-Bench.