Reinforcement Learning (RL) has significantly enhanced reasoning in Large Language Models. However, applying state-of-the-art RL methods like Group Relative Policy Optimization (GRPO) to video-based Multimodal LLMs (MLLMs) remains sub-optimal. We identify two key exploration bottlenecks: (1) Advantage Vanishing: generating multiple responses from invariant video contexts leads to near-uniform rewards, neutralizing learning signals after in-group standardization; (2) Visual Overfitting: standard RL lacks mechanisms to regularize sensitivity to visual perturbations under single-epoch training, causing policies to overfit to sparse temporal cues. To address these, we propose Dual-Group Exploration, a framework revitalizing learning signals at both input and output levels. At the input level, a Dual-Group Sampling strategy breaks visual determinism via complementary temporal sequences. At the output level, an Entropy-rated ELO Advantage captures inter-group quality variances. Concurrently, a Cross-Visual Counterfactual Weight establishes a Visual Consistency Region, safely driving exploration by enforcing reasoning consistency across varied views. Achieving state-of-the-art performance across six video reasoning benchmarks, our method demonstrates that consistency-guided exploration is beneficial for video understanding.

Mixture-of-experts variants of parameter-efficient fine-tuning enable per-token specialization, but they introduce additional trainable routers and expert parameters, increasing memory and training costs. This undermines the core goal of parameter efficient finetuning. We propose Monkey Jump, which brings MoE-style specialization to PEFT without adding extra trainable parameters for experts and routers. Instead of introducing new PEFT adapters as experts, Monkey Jump treats the PEFT adapters already present in each Transformer block (e.g., query, key, value, up, and down projections) as implicit experts and routes tokens among them. Routing is performed via k-means clustering with EMA-updated centers—no gradients, no learned parameters. We theoretically show that token-wise routing increases expressivity and can outperform shared adapters by avoiding cancellation effects. In multi-task experiments spanning 14 text, 14 image, and 19 video benchmarks, Monkey Jump achieves competitive performance with MoE-PEFT methods while using 5.3–9.2× fewer trainable parameters than the MoE variant of the same base method, up to 48% lower memory, and 1.5–2× faster training. Monkey Jump is architecture-agnostic and can be applied to any adapter-based PEFT method.

This paper argues that generating output tokens is more effective than using pooled representations for prediction tasks because token-level generation retains more mutual information. Since LLMs are trained on massive text corpora using next-token prediction, generation aligns naturally with their learned behavior. Using the Data Processing Inequality (DPI), we provide both theoretical and empirical evidence supporting this claim. However, autoregressive models face two key challenges when used for prediction: (1) exposure bias, where the model sees ground-truth tokens during training but relies on its own predictions during inference, leading to errors, and (2) format mismatch, where discrete tokens do not always align with the task’s required output structure. To address these challenges, we introduce PredGen (Predicting Through Generating), an end-to-end framework that (i) uses scheduled sampling to reduce exposure bias, and (ii) introduces a task adapter to convert the generated tokens into structured outputs. Additionally, we introduce Writer-Director Alignment Loss (WDAL), which ensures consistency between token generation and final task predictions, improving both text coherence and numerical accuracy. We evaluate PredGen on multiple classification and regression benchmarks. Our results show that PredGen consistently outperforms standard baselines, demonstrating its effectiveness in structured prediction tasks.

We propose Row-Column Fine-Tuning(RoCoFT), a parameter-efficient fine-tuning method for large language models based on updating only a few rows and columns of the weight matrices in transformers. Through extensive experiments with medium-sized LMs like RoBERTa and DeBERTa, and larger LMs like Bloom-7B, Llama2-7B, and Llama2-13B, we show that our method gives comparable or better accuracies than state-of-the-art Parameter-Efficient Finetuning methods while also being more memory and computation-efficient. We also study the reason behind the effectiveness of our method with tools from neural tangent kernel theory. We empirically demonstrate that our kernel, constructed using a restricted set of row and column parameters, is numerically close to the full-parameter kernel and gives comparable classification performance. Ablation studies are conducted to investigate the impact of different algorithmic choices, including the robustness of RoCoFT to any selection of rows and columns, as well as the optimal rank for the effective implementation of our method.