Latent Visual Reasoning (LVR) unfolds reasoning in a continuous visual embedding space, markedly improving Vision-Language Models (VLMs) on vision-intensive tasks. However, our experiments reveal a hidden cost: LVR inevitably undermines VLMs’ strong textual reasoning ability, causing substantial degradation on logic-intensive tasks. This motivates us to unify both reasoning modes within a single VLM that adaptively selects the appropriate one. We propose MoLaT (Mixture of Latent and Textual Reasoning), a novel method that trains VLMs via reinforcement learning (RL) to select LVR and text-only reasoning adaptively. MoLaT is trained in two stages. Supervised fine-tuning endows the VLM with a dual-mode capability using mixed CoT data. Then, MoLaT introduces two complementary designs: 1) The mode-adaptive advantage decouples the mode-selection signal from per-rollout correctness reward at the token level: mode-selection tokens receive dedicated cross-mode gradients, while content tokens retain the group-relative advantage. 2) The entropy-modulated advantage weight uses sequence-level entropy as a difficulty proxy, modulating signal strength to prevent premature mode commitment. Experiments confirm that MoLaT improves both vision-intensive and logic-intensive abilities, validating adaptive mode selection as key to general-purpose reasoning VLMs.

Real-world deployments of large language models (LLMs) increasingly rely on user-provided instructions that are long, precise, and tailored to specific workflows. In such settings, aggregate leaderboard performance is often a weak proxy for a model’s ability to faithfully execute the particular instruction that a practitioner intends to deploy. To address this gap, we introduce FlexBench/FlexEval, an instruction-conditioned evaluation framework that instantiates a specialized test suite from a user-provided target instruction. Given a target instruction, FlexBench (i) derives a set of verifiable evaluation dimensions and (ii) generates an instruction-conditioned conversation corpus. FlexEval then aggregates dimension-level judgments into two complementary metrics: Coverage, which measures workflow progress, and Achievement, which measures conditional correctness. We evaluate the framework on nine multi-turn target instructions spanning diverse scenario categories and workflow topologies, across nine contemporary LLMs over 3,888 generated dialogues. The results show that model rankings and failure modes vary substantially across target instructions, highlighting the need for instruction-conditioned evaluation beyond fixed aggregate benchmarks. Our source code is publicly available at https://github.com/Coryyy/FlexBench.