It's Not Always Sycophancy: Measuring LLM Conformity as a Function of Epistemic Uncertainty

Large language models (LLMs) are known to abandon their initial stance to conform to user pushback. While prior research largely attributes this behavior to sycophancy learned during reinforcement learning from human feedback, we hypothesize that conformity is also driven by a model's epistemic uncertainty at inference time. In this paper, we introduce MUSE, a two-stage evaluation framework to disentangle the mechanisms driving LLM conformity. Specifically, MUSE maps a model's epistemic uncertainty in responding to a query against its likelihood to yield to user pushback in a subsequent turn. We demonstrate that the mechanisms driving conformity extend beyond sycophancy alone. Specifically, we characterize two distinct factors that jointly drive conformity: sycophantic conformity, where a model aligns with user pushback even with absolute certainty in its initial response, and uncertainty-driven conformity, where a model's likelihood for conformity increases alongside its uncertainty. Furthermore, we conduct ablation studies to demonstrate that both sycophantic conformity and uncertainty-driven conformity grow with 1) the LLM's perceived expertise of the user and 2) the plausibility of the user's suggestions. More broadly, MUSE informs more targeted intervention strategies by distinguishing alignment-induced sycophancy and training-corpora-driven uncertainty.

StepAudio 2.5 Technical Report

Unified audio-language modeling has emerged as a prominent trend in modern speech systems, promising to bring the reasoning capabilities of large language models to auditory tasks. However, existing unified foundations often struggle to match the depth of specialized systems across automatic speech recognition (ASR), text-to-speech synthesis (TTS), and realtime spoken interaction. Bridging this gap remains an open challenge. This report presents StepAudio 2.5, a unified audio-language foundation model that matches or exceeds specialized systems across all three capabilities. Rather than treating these tasks as architecturally distinct, we operate on the premise that once text and audio share a multimodal representational space, task specialization becomes a matter of operational regimes: data construction, optimization targets, and decoding constraints. Guided by this insight, we advance the post-training paradigm from standard supervised learning to task-tailored Reinforcement Learning from Human Feedback (RLHF), using it as the primary mechanism to define complex optimization targets. We leverage this RLHF-centric alignment, alongside specialized decoding, to shape a shared backbone into three distinct operational modes. Concretely, the ASR branch advances transcription efficiency via verifiable multi-token decoding; the TTS branch achieves controllable, expressive synthesis through preference-based RLHF and context-rich supervision; and the Realtime branch realizes low-latency, persona-consistent dialogue via generative reward modeling within an RLHF framework. On standard benchmarks, StepAudio 2.5 achieves state-of-the-art results across ASR, TTS, and Realtime, demonstrating that a singular audio-language foundation can successfully internalize the distinct deployment objectives of speech understanding, generation, and live interaction.

DuplexSLA: A Full-Duplex Spoken Language Model with Synchronized Speech, Language, and Action

Recent advances in spoken dialogue language models have shifted from turn-based to full-duplex designs, where the model continuously listens to the user while generating responses. However, existing duplex backbones still lack a native channel for in-conversation planning and tool calling, leaving real-time agentic behaviour either tied to turn boundaries or relegated to an external cascade. We propose DuplexSLA, a native full-duplex Speech-Language-Action foundation model that decodes assistant audio together with a structured action stream on a shared 160 ms chunk timeline. DuplexSLA is built on a dual-stream three-channel formulation: a continuous user audio channel, a discrete assistant audio channel, and a rate-limited textual action channel, all decoded jointly by a single backbone, so that listening, speaking, planning, and tool calling unfold on one shared clock. Two capabilities define the model: (1) semantic-driven turn-taking control, where interruption, pause, and backchannel are handled inside the same backbone instead of by an external semantic VAD; and (2) in-conversation planning and tool calling, where planning text and structured tool calls are emitted on the action channel without halting assistant audio, so that multi-action and backchannel-triggered tool use are interleaved with ongoing speech. To evaluate these capabilities together, we further construct DuplexSLA-Bench, a duplex benchmark covering pause, interrupt, and backchannel turn-taking together with three styles of in-conversation tool calling. Our project page, interactive demos, and the DuplexSLA-Bench evaluation suite are publicly available at https://github.com/hyzhang24/DuplexSLA.

GaussFly: Contrastive Reinforcement Learning for Visuomotor Policies in 3D Gaussian Fields

Learning visuomotor policies for Autonomous Aerial Vehicles (AAVs) relying solely on monocular vision is an attractive yet highly challenging paradigm. Existing end-to-end learning approaches directly map high-dimensional RGB observations to action commands, which frequently suffer from low sample efficiency and severe sim-to-real gaps due to the visual discrepancy between simulation and physical domains. To address these long-standing challenges, we propose GaussFly, a novel framework that explicitly decouples representation learning from policy optimization through a cohesive real-to-sim-to-real paradigm. First, to achieve a high-fidelity real-to-sim transition, we reconstruct training scenes using 3D Gaussian Splatting (3DGS) augmented with explicit geometric constraints. Second, to ensure robust sim-to-real transfer, we leverage these photorealistic simulated environments and employ contrastive representation learning to extract compact, noise-resilient latent features from the rendered RGB images. By utilizing this pre-trained encoder to provide low-dimensional feature inputs, the computational burden on the visuomotor policy is significantly reduced while its resistance against visual noise is inherently enhanced. Extensive experiments in simulated and real-world environments demonstrate that GaussFly achieves superior sample efficiency and asymptotic performance compared to baselines. Crucially, it enables robust and zero-shot policy transfer to unseen real-world environments with complex textures, effectively bridging the sim-to-real gap.

Stop Listening to Me! How Multi-turn Conversations Can Degrade Diagnostic Reasoning

Patients and clinicians are increasingly using chatbots powered by large language models (LLMs) for healthcare inquiries. While state-of-the-art LLMs exhibit high performance on static diagnostic reasoning benchmarks, their efficacy across multi-turn conversations, which better reflect real-world usage, has been understudied. In this paper, we evaluate 17 LLMs across three clinical datasets to investigate how partitioning the decision-space into multiple simpler turns of conversation influences their diagnostic reasoning. Specifically, we develop a "stick-or-switch" evaluation framework to measure model conviction (i.e., defending a correct diagnosis or safe abstention against incorrect suggestions) and flexibility (i.e., recognizing a correct suggestion when it is introduced) across conversations. Our experiments reveal the conversation tax, where multi-turn interactions consistently degrade performance when compared to single-shot baselines. Notably, models frequently abandon initial correct diagnoses and safe abstentions to align with incorrect user suggestions. Additionally, several models exhibit blind switching, failing to distinguish between signal and incorrect suggestions.

Learning When to Sample: Confidence-Aware Self-Consistency for Efficient LLM Chain-of-Thought Reasoning

Large language models (LLMs) achieve strong reasoning performance through chain-of-thought (CoT) reasoning, yet often generate unnecessarily long reasoning paths that incur high inference cost. Recent self-consistency-based approaches further improve accuracy but require sampling and aggregating multiple reasoning trajectories, leading to substantial additional computational overhead. This paper introduces a confidence-aware decision framework that analyzes a single completed reasoning trajectory to adaptively select between single-path and multi-path reasoning. The framework is trained using sentence-level numeric and linguistic features extracted from intermediate reasoning states in the MedQA dataset and generalizes effectively to MathQA, MedMCQA, and MMLU without additional fine-tuning. Experimental results show that the proposed method maintains accuracy comparable to multi-path baselines while using up to 80\% fewer tokens. These findings demonstrate that reasoning trajectories contain rich signals for uncertainty estimation, enabling a simple, transferable mechanism to balance accuracy and efficiency in LLM reasoning.

CausalWrap: Model-Agnostic Causal Constraint Wrappers for Tabular Synthetic Data

Tabular synthetic data generators are typically trained to match observational distributions, which can yield high conventional utility (e.g., column correlations, predictive accuracy) yet poor preservation of structural relations relevant to causal analysis and out-of-distribution (OOD) reasoning. When the downstream use of synthetic data involves causal reasoning -- estimating treatment effects, evaluating policies, or testing mediation pathways -- merely matching the observational distribution is insufficient: structural fidelity and treatment-mechanism preservation become essential. We propose CausalWrap (CW), a model-agnostic wrapper that injects partial causal knowledge (PCK) -- trusted edges, forbidden edges, and qualitative/monotonic constraints -- into any pretrained base generator (GAN, VAE, or diffusion model), without requiring access to its internals. CW learns a lightweight, differentiable post-hoc correction map applied to samples from the base generator, optimized with causal penalty terms under an augmented-Lagrangian schedule. We provide theoretical results connecting penalty-based optimization to constraint satisfaction and relating approximate factorization to joint distributional control. We validate CW on simulated structural causal models (SCMs) with known ground-truth interventions, semi-synthetic causal benchmarks (IHDP and an ACIC-style suite), and a real-world ICU cohort (MIMIC-IV) with expert-elicited partial graphs. CW improves causal fidelity across diverse base generators -- e.g., reducing average treatment effect (ATE) error by up to 63% on ACIC and lifting ATE agreement from 0.00 to 0.38 on the intensive care unit (ICU) cohort -- while largely retaining conventional utility.

PRISM: Differentially Private Synthetic Data with Structure-Aware Budget Allocation for Prediction

Differential privacy (DP) provides a mathematical guarantee limiting what an adversary can learn about any individual from released data. However, achieving this protection typically requires adding noise, and noise can accumulate when many statistics are measured. Existing DP synthetic data methods treat all features symmetrically, spreading noise uniformly even when the data will serve a specific prediction task. We develop a prediction-centric approach operating in three regimes depending on available structural knowledge. In the causal regime, when the causal parents of $Y$ are known and distribution shift is expected, we target the parents for robustness. In the graphical regime, when a Bayesian network structure is available and the distribution is stable, the Markov blanket of $Y$ provides a sufficient feature set for optimal prediction. In the predictive regime, when no structural knowledge exists, we select features via differentially private methods without claiming to recover causal or graphical structure. We formalize this as PRISM, a mechanism that (i) identifies a predictive feature subset according to the appropriate regime, (ii) constructs targeted summary statistics, (iii) allocates budget to minimize an upper bound on prediction error, and (iv) synthesizes data via graphical-model inference. We prove end-to-end privacy guarantees and risk bounds. Empirically, task-aware allocation improves prediction accuracy compared to generic synthesizers. Under distribution shift, targeting causal parents achieves AUC $\approx 0.73$ while correlation-based selection collapses to chance ($\approx 0.49$).

Risk-Equalized Differentially Private Synthetic Data: Protecting Outliers by Controlling Record-Level Influence

When synthetic data is released, some individuals are harder to protect than others. A patient with a rare disease combination or a transaction with unusual characteristics stands out from the crowd. Differential privacy provides worst-case guarantees, but empirical attacks -- particularly membership inference -- succeed far more often against such outliers, especially under moderate privacy budgets and with auxiliary information. This paper introduces risk-equalized DP synthesis, a framework that prioritizes protection for high-risk records by reducing their influence on the learned generator. The mechanism operates in two stages: first, a small privacy budget estimates each record's "outlierness"; second, a DP learning procedure weights each record inversely to its risk score. Under Gaussian mechanisms, a record's privacy loss is proportional to its influence on the output -- so deliberately shrinking outliers' contributions yields tighter per-instance privacy bounds for precisely those records that need them most. We prove end-to-end DP guarantees via composition and derive closed-form per-record bounds for the synthesis stage (the scoring stage adds a uniform per-record term). Experiments on simulated data with controlled outlier injection show that risk-weighting substantially reduces membership inference success against high-outlierness records; ablations confirm that targeting -- not random downweighting -- drives the improvement. On real-world benchmarks (Breast Cancer, Adult, German Credit), gains are dataset-dependent, highlighting the interplay between scorer quality and synthesis pipeline.

Private Prediction via Shrinkage

We study differentially private prediction introduced by Dwork and Feldman (COLT 2018): an algorithm receives one labeled sample set $S$ and then answers a stream of unlabeled queries while the output transcript remains $(\varepsilon,δ)$-differentially private with respect to $S$. Standard composition yields a $\sqrt{T}$ dependence for $T$ queries. We show that this dependence can be reduced to polylogarithmic in $T$ in streaming settings. For an oblivious online adversary and any concept class $\mathcal{C}$, we give a private predictor that answers $T$ queries with $|S|= \tilde{O}(VC(\mathcal{C})^{3.5}\log^{3.5}T)$ labeled examples. For an adaptive online adversary and halfspaces over $\mathbb{R}^d$, we obtain $|S|=\tilde{O}\left(d^{5.5}\log T\right)$.