Caliper: Probing Lexical Anchors versus Causal Structure in LLMs

Large language models reach 50 to 70% accuracy on causal reasoning benchmarks such as CLadder, but it is unclear whether this reflects structural reasoning or lexical pattern matching. We introduce Caliper, a controlled perturbation that replaces semantic variable names with placeholder tokens while preserving the causal graph and probabilistic specification of each question. Across nine instruction-tuned LLMs from 3.8B to 671B and three causal reasoning benchmarks, lexical anonymization yields robust accuracy drops of +7.6, +27.0, and +11.1 pp on a local 3.8B-14B set, rising to +29.6 and +18.0 pp on CRASS and e-CARE across nine frontier models spanning the 2024-2026 generations. Of 40 engaged model-by-benchmark cells, 39 show a positive gap, and the gap collapses by 17x on CLadder's pseudoword subset. Structured scaffolding and few-shot in-context learning each narrow the gap, but mainly by lowering P0 accuracy on smaller models rather than recovering P1. Current instruction-tuned LLMs, evaluated zero-shot, show little evidence of structural causal reasoning once lexical anchors are removed.

MARS: Multi-rate Aggregation of Recency Signals for Sequential Recommendation across Sparse and Dense Regimes

Sequential recommenders weight historical interactions either through positional self-attention as in Transformers or through a single implicit decay schedule as in State-Space Models. Neither makes the multi-scale temporal structure of real user behaviour explicit. We propose MARS, an encoder-agnostic aggregation operator that consumes real timestamps and produces K summaries emphasising distinct recency scales, fused by a context-adaptive gate. MARS adds at most 6% parameters and runs in $\mathcal{O}(LdK)$ time. MARS adapts to data density by automatically selecting between two encoder instantiations: MARS-T (Transformer) for sparse data and MARS-M (Mamba) for dense data, based on the average sequence length of the training set. On five public benchmarks against ten Transformer- and Mamba-based baselines under a unified RecBole protocol, MARS attains the best HR@10 on every benchmark, with mean relative gain +19.7% over the strongest content-only Transformer baseline on sparse data (reaching +36.2% on Games) and +3.2% HR@10 / +0.9% NDCG over SIGMA on dense ML-1M at 42% fewer MFLOPs, occupying the accuracy-efficiency Pareto frontier across the data-density spectrum. A backbone-only ablation isolates the marginal contribution of MARS at +4% to +19% HR@10 on sparse data and motivates the dual-instantiation design. The code is included in the supplementary material.

When Does Latent Reasoning Help? MeRa: Metric-Space Bias for Spatial Prediction

Latent reasoning has improved sequential recommendation by iteratively refining representations before prediction, but does it help spatial prediction? We find that the answer depends on whether reasoning is grounded in the underlying metric space. Without such grounding, latent reasoning degrades spatial prediction below the unmodified baseline, while a learned metric-space bias derived from pairwise distances produces consistent gains. We formalize this finding through MeRa (Metric-space Reasoning), a lightweight backbone-agnostic module that can be inserted between any sequence encoder and its prediction heads. On the GETNext backbone, the gap between reasoning without and with metric-space bias reaches 4.5% NDCG@10. MeRa achieves the best NDCG@10 on all three spatial prediction benchmarks among the compared methods, surpassing recent approaches such as GeoMamba and HMST. We prove that metric-space-constrained reasoning converges to a unique fixed point and that N-step reasoning is strictly more expressive than (N-1)-step reasoning. A controlled experiment on CLEVR with Euclidean distance confirms that the finding generalizes beyond geographic coordinates. The code is included in the supplementary material.

Ghost: Plausible Yet Unlearnable Trajectories via On-Manifold Substitution for Next-POI Privacy

A publisher who releases check-in trajectories inadvertently publishes a strong predictor of every user's future locations. We address this risk by generating unlearnable trajectories, perturbed sequences that yield victim models with degraded next-Point-of-Interest (next-POI) accuracy on clean test inputs. Direct ports of image-domain unlearnable examples fail on two counts. The published data must remain geographically and semantically plausible, and the perturbation must resist purification adversaries that exploit the structure of randomized defences. We propose Ghost, a manifold-aligned framework whose perturbations look like plausible human check-in sequences yet leave no learnable signal behind. Ghost steers each substitution onto the real-trajectory manifold through a frozen trajectory language model, so a denoising-bridge adversary has nothing to invert and a context-free frequency-table adversary recovers a near-uniform distribution. Across two standard benchmarks, and four attacker postures, Ghost achieves protection-gap competitive with the strongest deterministic baseline (PGD) while attaining the lowest restored accuracy under the bigram adaptive purification adversary on both datasets, and lies within one per-cell standard deviation of PGD on the protection-versus-purification-resistance plane. Ablations confirm the manifold prior subsumes the entropy-floor knob of prior randomized defences, with the frequency-table adversary's survival gap remaining within 0.04 even when twenty percent of the pairs are leaked.

FlowTime: Towards Continuous Generative Watch Time Prediction via Flow-based Personalized Priors

Watch time has emerged as a pivotal metric for optimizing deep user engagement in short-video recommender systems. However, current methods of watch time prediction (WTP) suffer from inherent paradigm-specific limitations. Direct Regression faces mean-collapse due to unimodal Gaussian assumptions, while Ordinal Regression is hampered by quantization errors from rigid discretization. Similarly, Discrete Generative Regression struggles with high inference latency and heuristic vocabulary design. Beyond these specific flaws, a shared deficiency is the inability to capture the intrinsic multimodality and heterogeneity of User-Item Interaction Patterns. To address these challenges, we first revisit the WTP problem from a causal perspective and identify these user-specific patterns as structural confounders that modulate watch time outcomes, where identical interests manifest as distinct watch time outcomes conditioned on diverse user habits. Then, we formally propose a new (or the fourth) paradigm -- Continuous Generative Regression, and introduce FlowTime, a novel method utilizing a One-step Generative Variational Autoencoder. FlowTime effectively circumvents the latency of iterative denoising while maintaining the expressivity of continuous latent spaces. Furthermore, we design a Flow-based Personalized Prior that leverages NFs to warp a standard Gaussian prior into a complex, history-conditioned manifold, thereby enabling the adaptive modeling of multimodal interaction patterns. Finally, we build TimeRec, the first open-source WTP Library, alongside a novel personalization metric to establish a rigorous benchmarking standard. Extensive offline experiments and online A/B tests demonstrate FlowTime's significant superiority over SOTA methods.

SteerFace: Debiasing Synthetic Face Generation via Adaptive Residue Perturbation

The shortage of legally compliant data for face recognition training has sparked growing interest in using synthetic data as an alternative. While recent diffusion-based methods enable the generation of photorealistic face images with strong identity adherence and data diversity, their downstream recognition performance still exhibits a significant synthetic-real gap. This paper identifies visual tendency as a previously underexplored limitation, whereby synthetic data exhibit an unrealistic prevalence of visual attributes and thus deviate from the real-data distribution. Visual tendency can be attributed to the generator's conditioning on identity embeddings, through which co-occurring residual visual cues are unintentionally absorbed into learned identity semantics. To discourage the generator from exploiting such visual cues, this paper proposes SteerFace, a simple and efficient training framework that perturbs identity embeddings by steering them toward random orthogonal directions on the embedding hypersphere. The perturbation serves as an identity-preserving regularizer that penalizes the generator's reliance on non-identity components, as supported by theoretical analysis. This paper further introduces an adaptive strategy that learns perturbation strengths with both sample-wise preference and favorable overall statistics. Extensive experiments show that SteerFace effectively mitigates visual tendency, outperforms prior methods in downstream face recognition, and generalizes well across different training datasets and generation pipelines.

TRACER: Turn-level Regret Matching with Inner Reinforcement Credit for Cooperative Multi-LLM Reasoning

Large language models increasingly rely on either reinforcement learning or multi-agent prompting to improve reasoning, yet these two paradigms remain difficult to combine. Directly applying single-agent reinforcement learning to multi-turn multi-agent systems faces following dilemmas: i) Sparse rewards, role-level free-riding and excessive training overhead. ii) Agents only imitate to collaborate. iii) Fixed collaboration protocol falls into oscillating local optimum. We introduce TRACER, a turn-level reinforcement framework for cooperative multi-LLM reasoning. TRACER separates collaborative decision making into a controller-regret layer, where controllers learn whether the agents should speak or skip the current round through regret matching, and a generation-credit layer, which optimizes proposer and reviewer utterances with role-specific GSPO rewards. This design i) assigns credit at the level of both action modes and generated utterances, thus avoiding free-riding and sparse rewards. We only expand the choices made by the controllers, thus greatly reducing computational cost of training. Moreover, ii) agents acquire collaborative capability as they learn when to utter and what to speak. Finally, iii) by designing binary actions ingeniously, we extend classical game theory established for finite action spaces to deep learning, thus achieving mathematically rigorous convergence. We train all local RL-style methods on the GSM8K training split and evaluate on held-out GSM8K, MATH500, and GPQA-Diamond to measure in-domain accuracy, cross-benchmark generalization, inference cost, and correction-preservation behavior. The resulting framework provides a compact and reproducible testbed for studying learned collaboration policies beyond fixed debate, voting, or aggregation protocols. Code is available at https://github.com/Shark-Forest/TRACER.

Disentangling Generation and Regression in Stochastic Interpolants for Controllable Image Restoration

Recent advances in Image Restoration (IR) have been largely driven by generative methods such as Diffusion Models and Flow Matching, which excel in synthesizing realistic textures while suffering from slow multi-step inference and compromised pixel fidelity. In contrast, classical regression-based IR methods excel precisely in these aspects, offering single-step efficiency and high pixel-level reconstruction fidelity. To bridge this gap, we propose DiSI, a unified framework that Disentangles the underlying Stochastic Interpolant process into independent generation and regression components. This decoupling endows DiSI with remarkable versatility, enabling a continuous and controllable transition from a pure regression process to a fully generative one. Technically, we instantiate this framework with two specific sampling trajectories, accompanied by a unified sampler for high-quality, few-step inference on arbitrary trajectories. Furthermore, we design a dual-branch U-Net style transformer network in pixel space, using a dedicated branch to enhance conditional guidance while ensuring high throughput. Extensive experiments demonstrate that DiSI efficiently achieves competitive results on various IR tasks, while uniquely offering the inference-time flexibility to control the distortion-perception trade-off within a single model.

Can Vision Models Truly Forget? Mirage: Representation-Level Certification of Visual Unlearning

Machine unlearning in Vertical Federated Learning (VFL) has attracted growing interest, yet existing methods certify forgetting solely using output-level metrics. We challenge these claims by introducing Mirage, a representation-level auditing framework comprising four complementary diagnostics: Linear Probe Recovery (LPR), Centered Kernel Alignment (CKA), Feature Separability Scoring, and Layer-Wise Recovery Analysis. Through experiments across seven datasets and seven baseline methods following recent VFL unlearning protocols, Mirage reveals three key findings: (i) Forgetting gap: methods that pass output-level certification still retain substantial class structure in their representations, with LPR exceeding the retrained baseline by up to 15.4 points; CKA shows these models remain structurally closer to the original than to the retrained reference, while separability scores indicate persistent geometric discrimination. (ii) Unlearning trilemma: no existing method simultaneously achieves high utility, output-level forgetting, and representation-level forgetting. (iii) Class-sample asymmetry: class-level forgetting leaves strong representational traces (LPR up to 97%), whereas sample-level forgetting is indistinguishable from chance (LPR approx. 50%); layer-wise analysis further shows residual class information persists across network depths. These findings call for representation-aware evaluation standards in federated unlearning research.

AdaMamba: Adaptive Frequency-Gated Mamba for Long-Term Time Series Forecasting

Accurate long-term time series forecasting (LTSF) requires the capture of complex long-range dependencies and dynamic periodic patterns. Recent advances in frequency-domain analysis offer a global perspective for uncovering temporal characteristics. However, real-world time series often exhibit pronounced cross-domain heterogeneity where variables that appear synchronized in the time domain can differ substantially in the frequency domain. Existing frequency-based LTSF methods often rely on implicit assumptions of cross-domain homogeneity, which limits their ability to adapt to such intricate variability. To effectively integrate frequency-domain analysis with temporal dependency learning, we propose AdaMamba, a novel framework that endogenizes adaptive and context-aware frequency analysis within the Mamba state-space update process. Specifically, AdaMamba introduces an interactive patch encoding module to capture inter-variable interaction dynamics. Then, we develop an adaptive frequency-gated state-space module that generates input-dependent frequency bases, and generalizes the conventional temporal forgetting gate into a unified time-frequency forgetting gate. This allows dynamic calibration of state transitions based on learned frequency-domain importance, while preserving Mamba's capability in modeling long-range dependencies. Extensive experiments on seven public LTSF benchmarks and two domain-specific datasets demonstrate that AdaMamba consistently outperforms state-of-the-art methods in forecasting accu racy while maintaining competitive computational efficiency. The code of AdaMamba is available at https://github.com/XDjiang25/AdaMamba.