Reasoning language models are controlled through explicit modes such as Think and No-think, yet we find that these behaviors are largely governed by a few token-level triggers rather than high-level instructions. Through attention analysis and controlled prompting experiments, we show that a leading “Okay” token induces reasoning behavior, while the newline pattern following ‘</think>‘ suppresses it. Based on this observation, we propose Mid-Think, a simple training-free prompting format that combines these triggers to achieve intermediate-budget reasoning, consistently outperforming fixed-token and prompt-based baselines in terms of the accuracy–length trade-off. Furthermore, applying Mid-Think to RL training after SFT reduces training time by approximately 15% while improving final performance of Qwen3-8B on AIME from 69.8% to 72.4% and on GPQA from 58.5% to 61.1%, demonstrating its effectiveness for both inference-time control and RL-based reasoning training.

Multimodal large language models (MLLMs) remain unreliable on spatial multiple-choice questions, and their failures are often attributed to poorly attended visual information. We identify a complementary failure mode, spatial lexical bias: a spatial relation word added to the answer options can act as a lexical-semantic distractor that draws the model’s decision toward that option. Using nine open-weight MLLMs, we show that this phenomenon is widespread. We then isolate diagnostic cases in which a model answers a binary spatial question correctly yet consistently chooses a newly added third spatial option, which we call binary-stable but ternary-fragile cases. Leveraging mechanistic interpretability tools on these cases, we find that the failure arises on the language side rather than the visual side: visual attention analyses and residual-stream probes show the correct spatial relation remains internally available, while irrelevant-option controls, activation patching, and sparse component interventions trace the bias to specific LLM-side channels and neurons. Accordingly, we show that a lightweight LLM-only DPO update on tiny single-object-pair synthetic data mitigates the bias, lifting four-way robust accuracy by up to 100 points on synthetic data, and by 68.0, 32.6, and 20.1 points on broader evaluation datasets WhatsUp, SpatialMQA-Direct, and VSR.

Hybrid thinking enables LLMs to switch between reasoning and direct answering, offering a balance between efficiency and reasoning capability. Yet our experiments reveal that current hybrid thinking LLMs only achieve partial mode separation: reasoning behaviors often leak into the no-think mode. To understand and mitigate this, we analyze the factors influencing controllability and identify four that matter most: (1) larger data scale, (2) using think and no-think answers from different questions rather than the same question, (3) a moderate increase in no-think data number, and (4) a two-phase strategy that first trains reasoning ability and then applies hybrid think training. Building on these findings, we propose a practical recipe that, compared to standard training, can maintain accuracy in both modes while significantly reducing no-think output length (from 1085 to 585 on MATH500) and occurrences of reasoning-supportive tokens such as "wait" (from 5917 to 522 on MATH500). Our findings highlight the limitations of current hybrid thinking and point the ways for enhancing its controllability.

In this paper, we present a novel pipeline for the XLLM Shared Task-III: Large Language Model for Structural Reasoning (LLM-SR). Our pipeline addresses key challenges in automatic process-reward training data construction, such as high manual annotation costs, limited accuracy of large models in structured data processing, and dependency on auxiliary information for validation. To overcome these limitations, we first decompose the construction process into extraction and validation phases. Leveraging model-generated annotations, we produce pseudo-labeled data and iteratively refine model performance. Second, by analyzing structured data patterns, we encode structural constraints into a rule-based module and fine-tune the model with Gradient Reward Policy Optimization (GRPO), significantly improving structured data extraction success rates. Finally, we train the model to generate critical responses that assess evidence-conclusion relationships, thus enhancing validation reliability. Experimental results demonstrate that our pipeline outperforms models with an order of magnitude more parameters and achieves the first position on the task.

Long-context capability is considered one of the most important abilities of LLMs, as a truly long context-capable LLM shall enable its users to effortlessly process many originally exhausting tasks — e.g., digesting a long-form document to find answers v.s., directly asking an LLM about it. However, existing real-task-based long-context evaluation benchmarks have a few major shortcomings. For instance, some Needle-in-a-Haystack-like benchmarks are too synthetic, and therefore do not represent the real world usage of LLMs. While some real-task-based benchmarks like LongBench avoid this problem, such benchmarks are often formed in a way where each data sample has a fixed sequence length, which not only makes them solely suitable for models with a certain range of context windows, but also lacks a proxy to know at what length the model/method-of-interest would fail. Last, most benchmarks tend to not provide proper metrics to separate long-context performance from the model’s baseline ability, so when conducting a cross-model/recipe comparison, such conflation makes the user unable to understand how exactly one model or recipe excels at the long-context task in relation to its baseline ability. To address these issues, we introduce a length-controllable, real-life reflective benchmark with a novel metric that disentangles baseline knowledge from long-context capabilities. Experiments demonstrate the superiority of our datasets in effectively evaluating LLMs. All assets are available at https://github.com/uservan/100-LongBench.git.