Multimodal large language models (MLLMs) are increasingly deployed as assistants that interact through text and images, making it crucial to evaluate contextual safety when risk depends on both the visual scene and the evolving dialogue. Existing contextual safety benchmarks are mostly single-turn and often miss how malicious intent can emerge gradually or how the same scene can support both benign and exploitative goals. We introduce the Multi-Turn Multimodal Contextual Safety Benchmark (MTMCS-Bench), a benchmark of realistic images and multi-turn conversations that evaluates contextual safety in MLLMs under two complementary settings, escalation-based risk and context-switch risk. MTMCS-Bench offers paired safe and unsafe dialogues with structured evaluation. It contains over 30 thousand multimodal (image+text) and unimodal (text-only) samples, with metrics that separately measure contextual intent recognition, safety-awareness on unsafe cases, and helpfulness on benign ones. Across eight open-source and seven proprietary MLLMs, we observe persistent trade-offs between contextual safety and utility, with models tending to either miss gradual risks or over-refuse benign dialogues. Finally, we evaluate five current guardrails and find that they mitigate some failures but do not fully resolve multi-turn contextual risks.

State-of-the-art (SOTA) reasoning post-training pipelines increasingly combine supervised fine-tuning (SFT) with reinforcement learning (RL) to balance external knowledge injection and self-exploration. In this work, we identify a critical but underexplored failure mode of these SFT+RL pipelines: catastrophic forgetting. To analyze the cause, we reveal a fundamental asymmetry between SFT and RL. SFT typically induces much larger, redundant parameter updates than RL, while RL’s updates are smaller and more parsimonious. Therefore, SFT updates can easily overwrite RL’s small non-robust updates, causing the forgetting of the knowledge learned by RL. To address this issue, we propose MIFO (MItigating FOrgetting between SFT and RL), a plug-and-play framework that integrates SFT into RL while explicitly constraining SFT-induced updates. MIFO mitigates forgetting through two complementary designs: (i) Data selection, which interleaves RL and SFT by selecting challenging samples based on rollout accuracy and optimizing only high-entropy tokens for SFT; and (ii) parameter freezing, which identifies RL-critical parameters and temporarily freezes them during SFT to prevent overwrite. Beyond stability, MIFO is data- and response-efficient and remains agnostic to the underlying SFT or RL algorithm. Empirically, MIFO achieves SOTA reasoning performance using only 1.5% of the SFT data and 20.4% of the RL data required by prior SOTA methods, while producing reasoning traces up to 10× shorter than the standard SFT→RL pipeline.

Multimodal Large Language Models (MLLMs) are increasingly deployed in real-world applications, yet their ability to make context-aware safety decisions remains limited. Existing methods often fail to balance oversensitivity (unjustified refusals of benign queries) and undersensitivity (missed detection of visually grounded risks), leaving a persistent gap in safety alignment. To address this issue, we introduce Safety-aware Contrastive Decoding (SafeCoDe), a real-time decoding framework that dynamically adjusts token generation based on multimodal context. SafeCoDe operates in two stages: (1) a contrastive decoding mechanism that highlights tokens sensitive to visual context by contrasting real and Gaussian-noised images, and (2) a global-aware token modulation strategy that integrates scene-level reasoning with token-level adjustment to adapt refusals according to the predicted safety verdict. Extensive experiments across diverse MLLM architectures and safety benchmarks, covering undersensitivity, oversensitivity, and general safety evaluations, show that SafeCoDe consistently improves context-sensitive refusal behaviors while preserving model helpfulness.

Curriculum learning (CL), which orders training data from easy to hard, has become a popular strategy for improving reasoning in large language models (LLMs). Yet prior work employs disparate difficulty metrics and training setups, leaving open fundamental questions: When does curriculum help? Which direction—forward or reverse—is better? And does the answer depend on what we measure? We address these questions through a unified offline evaluation framework that decomposes curriculum difficulty into five complementary dimensions: Problem Difficulty, Model Surprisal, Confidence Margin, Predictive Uncertainty, and Decision Variability. Through controlled post-training experiments on mathematical reasoning benchmarks with Llama3.1-8B, Mistral-7B, and Gemma3-4B, we find that: (i) no curriculum strategy dominates universally—the relative effectiveness of forward versus reverse CL depends jointly on model capability and task complexity; (ii) even within a single metric, samples at different difficulty levels produce distinct gains depending on task demands; and (iii) Task-aligned curricula focus on shaping the model’s final representations and generalization, whereas inner-state curricula modulate internal states such as confidence and uncertainty. Our findings challenge the notion of a universal curriculum strategy and offer actionable guidance across model and task regimes, with some metrics indicating that prioritizing decision-uncertain samples can further enhance learning outcomes.

Reinforcement learning (RL) is central to post-training, particularly for agentic models that require specialized reasoning behaviors. In this setting, model merging offers a practical mechanism for integrating multiple RL-trained agents from different tasks into a single generalist model. However, existing merging methods are designed for supervised fine-tuning (SFT), and they are suboptimal to preserve task-specific capabilities on RL-trained agentic models. The root is a task-vector mismatch between RL and SFT: on-policy RL induces task vectors that are highly sparse and heterogeneous, whereas SFT-style merging implicitly assumes dense and globally comparable task vectors. When standard global averaging is applied under this mismatch, RL’s non-overlapping task vectors that encode critical task-specific behaviors are reduced and parameter updates are diluted. To address this issue, we propose Reinforced Agent Merging (RAM), a distribution-aware merging framework explicitly designed for RL-trained agentic models. RAM disentangles shared and task-specific unique parameter updates, averaging shared components while selectively preserving and rescaling unique ones to counteract parameter update dilution. Experiments across multiple agent domains and model architectures demonstrate that RAM not only surpasses merging baselines, but also unlocks synergistic potential among agents to achieve performance superior to that of specialized agents in their domains.

Large Language Models (LLMs) rely heavily on large-scale reasoning data, but as such data becomes increasingly scarce, model self-improvement offers a promising alternative. However, this process can lead to model collapse, as the model’s output becomes overly deterministic with reduced diversity. In this work, we identify a new risk beyond model collapse, which we term the Superficial Self-Improved Reasoners phenomenon. This phenomenon indicates that while self-improvement enhances in-domain (ID) reasoning accuracy, it degrades the model’s generalized reasoning capability on out-of-domain (OOD) datasets, as the model tends to memorize the training data. Our analyses of layer importance and parameter changes reveal that reasoning-critical layers receive fewer updates compared to less relevant layers during self-improvement. To address this, we propose Iterative Model Merging (IMM), which balances reasoning improvements and generalization by merging the weights of the original and self-improved models. IMM effectively mitigates model collapse and improves generalized reasoning capability. Code is available at https://github.com/xiangchi-yuan/merge_syn

Language models (LMs) require effective episodic grounding—the ability to learn from and apply past experiences—to perform well at physical planning tasks. While current approaches struggle with scalability and integration of episodic memory, which is particularly limited for medium-sized LMs (7B parameters), larger LMs (70-405B) offer untapped potential through their hierarchical representations and extensive pre-trained knowledge. Therefore, to unlock larger LMs’ potential for grounding, we present a scalable weak-to-strong episodic learning framework that efficiently transfers episodic behaviors from smaller to larger LMs. It uses Monte Carlo tree search for structured experience collection with a novel distillation method that preserves LM capabilities while incorporating episodic memory. This enables larger LMs to leverage their inherent advantages for improved physical planning. Experiments show our solution outperforms top proprietary LMs by 3.45% across diverse planning and question-answering tasks. Layer-wise probing reveals systematic improvements in task alignment, particularly in later LM layers. It shows stable generalization to even unseen scenarios, even as planning steps increase, whereas baselines deteriorate sharply beyond a complexity threshold of four planning steps.

Generative models such as Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) trained on massive datasets can lead them to memorize and inadvertently reveal sensitive information, raising ethical and privacy concerns. While some prior works have explored this issue in the context of LLMs, it presents a unique challenge for MLLMs due to the entangled nature of knowledge across modalities, making comprehensive unlearning more difficult. To address this challenge, we propose Modality Aware Neuron Unlearning (MANU), a novel unlearning framework for MLLMs designed to selectively clip neurons based on their relative importance to the targeted forget data, curated for different modalities. Specifically, MANU consists of two stages: important neuron selection and selective pruning. The first stage identifies and collects the most influential neurons across modalities relative to the targeted forget knowledge, while the second stage is dedicated to pruning those selected neurons. MANU effectively isolates and removes the neurons that contribute most to the forget data within each modality, while preserving the integrity of retained knowledge. Our experiments conducted across various MLLM architectures illustrate that MANU can achieve a more balanced and comprehensive unlearning in each modality without largely affecting the overall model utility.