Temporal reasoning benchmarks for Video-LLMs are often mediated by language, leaving room for linguistic shortcuts from option wording, answer correlations, or language priors. To reduce such shortcuts, we introduce TempCloze, a video cloze benchmark for evaluating visual temporal reasoning in Video-LLMs. Given the beginning and ending clips of a video, models must identify the true missing middle from four candidates. TempCloze contains 1,521 carefully filtered videos from seven sources, mainly long-take and egocentric videos. We construct same-source distractors along three dimensions: Semantic asks what event should happen, Alignment probes when it should occur, and Progression tests how it should unfold, while shared scenes and objects reduce appearance cues. Our evaluation of 10 proprietary and 21 open-source Video-LLMs reveals Alignment as the primary bottleneck: models often recognize plausible semantic content and local event progression but struggle with temporal alignment. We further conduct error pattern and behavioral sensitivity analyses on TempCloze-Mixed and TempCloze-Hard with four representative models to examine where errors arise and how candidate order, context direction, visible span, frame density, and test-time scaling influence model choices.

Reasoning in Large Language Models (LLMs) is primarily conducted in a discrete token space. Although recent approaches employing continuous concept tokens facilitate abstract reasoning, they introduce significant stability challenges. The accumulation of Out-of-Distribution (OOD) states frequently causes manifold deviation, while the absence of clear boundaries leads to indeterminate termination, often referred to as the anchoring problem. To address these limitations, we propose HybridThink. We formulate mode selection in LLM reasoning as a sequential decision process. A lightweight gating policy is trained with RL to decide at each step whether to propagate a continuous concept token or generate a discrete text token to anchor the semantic trajectory.This learned interleaving limits manifold deviation and resolves the fragility of static rules. Furthermore, we ensure stable convergence through a training strategy that incorporates Error-Centric Experience Replay (ECER) and Inductive Exploration Shaping (IES), prioritizing learning from failure modes, specifically drifting and premature termination. Empirical evaluations on mathematics and coding benchmarks demonstrate that HybridThink outperforms standard CoT and Soft Thinking baselines, achieving superior accuracy with reduced token consumption.

Continuous-space reasoning methods replace discrete intermediate tokens with probability-weighted soft embeddings, aiming to preserve semantic information that greedy Chain-of-Thought decoding discards. We show that this promise is limited by local representation geometry and the latent semantic manifold. When the next-token distribution is low-entropy and concentrated, a soft token becomes nearly collinear with the Top-1 embedding and collapses toward greedy decoding. When the distribution is high-entropy, diffuse, or stochastically perturbed, the continuous state may leave the greedy path but drift away from the local scale of naturally occurring candidate representations. We diagnose these two regimes with step-level probes over representation divergence, local norm calibration, and behavior-cloud membership. To address them, we propose Manifold Thinking, a training-free inference framework that combines structured tangent movement away from the Top-1 direction with local norm-calibrated projection back to the candidate-token neighborhood. Across mathematical and programming benchmarks, Manifold Thinking improves over discrete CoT baselines and baselines covering the latest stochastic Soft Thinking strategies, suggesting that useful continuous reasoning requires geometric correction rather than unconstrained linear token mixing alone.