All papers

The paper enables multimodal models to proactively ask clarifying questions when ambiguous inputs carry cross-modal safety risks.

Explores methods to extract the underlying decisive sub-questions that drive language model decisions across diverse generated outputs.

The paper studies how repeated input measurements can degrade out-of-sample prediction quality in tabular foundation models.

Introduces an auditing framework that checks counterfactual execution traces to ensure language model routing uses only legitimate information.

The paper introduces a framework to recalibrate, audit, and safely deploy confidence scores for language model outputs.

Analyzes how continual learning methods affect demographic fairness over time across eleven deepfake generation techniques.

The paper demonstrates vulnerabilities in unilateral weight obfuscation schemes used for on-device secure language model inference.

The paper proposes a selective safety framework for vision-language models using counterfactual pairs and representation density estimation.

The paper introduces an auditing framework that identifies failure steps in multi-agent systems by evaluating how information is utilized across handoffs.

The authors analyze how different optimizers leave distinct implicit bias fingerprints that influence training data reconstruction from model checkpoints.

This paper investigates attention heads responsible for contextual entrainment and evaluates whether their ablation transfers across different semantic relations.

This paper presents a post-hoc method that refines fine-tuned models by selectively reverting parameter updates along principal directions of base weights.

The authors propose a probe-based method to measure and calibrate systematic scoring biases in black-box LLM evaluation judges.

LR-Guard distills explicit safety reasoning into compact latent representations to speed up guardrail classification while enabling on-demand trace reconstruction.

The authors propose verifiable agent insurance using smart-contract escrow to financially compensate users when agents violate programmatically logged behavioral promises.

This paper proposes an inference-time intervention method to detect and correct modality misrouting that causes multimodal hallucinations.

This paper presents a post-hoc multicalibration framework to correct systematic biases such as position and model family in LLM evaluations.

This paper presents an evaluation benchmark and guardrail framework for detecting safety risks in multi-turn interactions with AI companions.

This paper proposes decoupling error proposing from error validation to filter invalid candidates under budget constraints for expert human review.

This paper introduces an evaluation framework and reward models to assess how well search agents ground generated answers in retrieved, independent sources.

This survey analyzes how verifier noise affects reinforcement learning with verifiable rewards across 58 studies and summarizes practical mitigation techniques.

This paper presents a compact decision model designed to screen malicious agent tool skills with high recall directly from specification files.

Evaluates SHAP attribution methods on sparse autoencoder features, finding weak correspondence with direct causal feature intervention in language models.

Presents ADMA, an adaptive deliberation controller that dynamically adjusts multi-agent critique and coordination to improve reasoning accuracy.

The paper introduces local contract-conditioned rewriting to prevent sensitive data leakage in email-generating language model agents cost-effectively.

The authors examine why language model agents fail by analyzing discrepancies between internal representations of state values and actual action choices.

The paper develops a sequential betting approach to reliably detect data distribution shifts in unstructured text and image streams over time.

The paper presents a unified node-centric framework to quantify and manage multi-agent system lifecycles across design, runtime, and post-hoc attribution.

The paper proposes a training-free framework that co-evolves an executable verifier program and harness state to self-improve agents on open-ended tasks.

Proposes a training-free framework that mitigates vision-language model hallucinations by admitting claims only when supported relative to a visual-null input.

Formulates machine unlearning for LLM agents across parametric memory, external persistent storage, and tool-accessible environments.

The paper applies an information-flow modeling methodology to assess public-sector AI recruitment tools under European Union AI Act requirements.

The paper presents a benchmark to evaluate whether formal specifications generated by language models accurately reflect natural-language requirements.

Proposes a decentralized identity framework using self-sovereign identity principles to establish verification and accountability for autonomous agents.

Routes gradients from imperfect monitor flags into sacrificial parameters to mitigate reward hacking during reinforcement learning.

Generates prompt-adapted adversarial images that block unauthorized visual questions on vision-language models while still answering legitimate prompts correctly.

Monitors model explanation drift on evolving graphs to provide early warnings before predictive uncertainty guarantees fail.

The paper presents StableJudge, an external lightweight sentry model that sanitizes prompt-injection attacks before LLM-as-a-judge scoring.

The paper proposes evaluating partial generation outputs to stop harmful model interactions early while avoiding unnecessary prompt rejections.

The paper presents a runtime auditor that predicts execution failures in multi-agent systems using loss-guided perturbations to identify hard failure cases.

The paper proposes an ethical governance scaffold based on reflective equilibrium to resolve conflicting instructions and policy rules in agentic workflows.

This paper proposes control-inspired gates to decide when language model agents should adapt their prompts without absorbing noisy user feedback.

The paper presents a benchmark evaluating how language model agents handle unauthorized tool call retries after receiving timeouts.

The paper studies whether language models can track temporal validity changes in parametric memory via adapter conditioning and clocks.

The paper examines how hypernetwork adapters resolve knowledge conflicts and compose information during continual learning in language models.

The paper proposes a multi-layer detector ensemble to identify deceptive model outputs while calibrating score uncertainty and false-positive rates.

Evaluates commercial language models against Universal Declaration of Human Rights principles, revealing misalignments driven by efficiency arguments.

Critiques trust calibration in human-AI decision making for open-ended problems and outlines dimensions for higher-order system alignment.

Proposes an online learning algorithm that decides whether language models should answer queries directly or defer to congested human experts.

This paper evaluates the capabilities and limitations of using activation steering vectors as an agent memory mechanism.

This paper introduces a benchmark that uses agent skill specifications to evaluate the reliability and procedural correctness of AI agents.

The paper proposes an execution-gating method to safely launch and cancel tool calls early during live speech recognition before user self-corrections.

The paper analyzes training instability cascades in large recommender systems and proposes targeted gradient, weight, and loss interventions.

This paper provides a systematization of knowledge on membership inference attacks in tabular machine learning across small-to-medium datasets.

The paper investigates how conditional flow models learn and potentially hallucinate non-additive interactions in scientific prediction tasks.

The paper proposes a latent-space unlearning method that preserves model performance on benign tasks conceptually adjacent to forgotten domains.

The paper introduces a reasoning-guided translation framework to scale multilingual safety and reasoning benchmarks while preserving original task meanings.

The paper uses in-context neurofeedback sessions to evaluate whether language models can intentionally control their internal activations.

The paper proposes a synthetic data pipeline that extracts atomic facts from source texts to guide factual pretraining for language models.

The paper detects backdoor triggers in multimodal language models by tracking output evolution trajectories across internal hidden layers.