RLCSD: Reinforcement Learning with Contrastive On-Policy Self-Distillation

On-policy self-distillation (OPSD) provides dense, token-level supervision for reasoning models by aligning a model's own distribution with the distribution it produces under privileged context, typically a verified solution. However, we show that the learning signal drawn from this distributional gap concentrates on style tokens rather than task-bearing ones, as the hinted model tends to produce more direct, shorter outputs. We term this pathology \emph{privilege-induced style drift}, which destabilizes training or causes response length to shrink. To address this, we propose \textbf{RLCSD} (Reinforcement Learning with Contrastive on-policy Self-Distillation), which mitigates this drift by contrasting the teacher-student gap under a correct hint against that under a wrong hint, suppressing the style shift that conditioning on a hint tends to induce regardless of correctness, and yielding a signal that is more concentrated on task-bearing tokens. Experiments on Qwen3 (1.7B/4B/8B) and Olmo-3-7B-Think across mathematical and logical reasoning show that RLCSD consistently outperforms GRPO and prior OPSD methods. We further show that the contrastive principle is general: it plugs into existing OPSD methods to improve them, and its underlying insight extends to the broader cross-model on-policy distillation setting.

VaaWIT: Visual-Aware Adaptation of Large Language Models for Multilingual Web Image Translation

Translating text embedded in Web images is crucial for improving content accessibility and cross-lingual information retrieval, particularly within social media and e-commerce domains. Although Large Vision-Language Models (LVLMs) have advanced multimodal understanding, applying them to Web image translation remains challenging due to the visual representation gap: standard encoders often prioritize high-level semantics over the fine-grained visual details required for recognizing diverse character morphologies. To address this challenge, we propose VaaWIT, an end-to-end framework that adapts Large Language Models for multilingual Web image translation. The framework introduces two key technical contributions: (1) a Dual-Stream Attention Module (DSAM), which facilitates bidirectional interaction between multilingual semantic features and detailed visual representations, thereby synthesizing unified features robust to textual variations; and (2) a Visual-Aware Adapter (VAA), a parameter-efficient fine-tuning strategy that dynamically injects these fused visual cues into the frozen LLM backbone. This design enables the model to align the visual context with linguistic reasoning effectively while minimizing computational costs. Extensive experiments on eight tasks on three public benchmarks demonstrate that VaaWIT significantly outperforms state-of-the-art (SOTA) open-source baselines and achieves competitive performance against proprietary models. These results validate the efficacy of integrating fine-grained visual perception into LLMs for complex Web content analysis.

Do We Really Need External Tools to Mitigate Hallucinations? SIRA: Shared-Prefix Internal Reconstruction of Attribution

Large vision-language models (LVLMs) often hallucinate when language priors dominate weak or ambiguous visual evidence. Existing contrastive decoding methods mitigate this problem by comparing predictions from the original image with those from externally perturbed visual inputs, but such references can introduce off-manifold artifacts and require costly extra forward passes. We propose SIRA, a training-free internal contrastive decoding framework that constructs a counterfactual reference inside the same LVLM by exploiting the staged information flow of multimodal transformers. Instead of removing visual information from the input, SIRA first lets image and text tokens interact through a shared prefix, forming an aligned multimodal state that preserves prompt interpretation, decoding history, positional structure, and early visual grounding. It then forks a counterfactual branch in later transformer layers, where attention to image-token positions is masked. This branch retains the shared multimodal context but lacks continued access to fine-grained visual evidence, yielding a language-prior-dominated internal reference for token-level contrast. During decoding, SIRA suppresses tokens that remain strong without late visual access and favors predictions whose advantage depends on the full visual pathway. Experiments on POPE, CHAIR, and AMBER with Qwen2.5-VL and LLaVA-v1.5 show that SIRA consistently reduces hallucinations while preserving descriptive coverage and incurring lower overhead than two-pass contrastive decoding. SIRA requires no training, external verifier, or perturbed input, and applies to open-weight LVLMs with white-box inference access.

Student-in-the-Loop Chain-of-Thought Distillation via Generation-Time Selection

Large reasoning models achieve strong performance on complex tasks through long chain-of-thought (CoT) trajectories, but directly transferring such reasoning processes to smaller models remains challenging. A key difficulty is that not all teacher-generated reasoning trajectories are suitable for student learning. Existing approaches typically rely on post-hoc filtering, selecting trajectories after full generation based on heuristic criteria. However, such methods cannot control the generation process itself and may still produce reasoning paths that lie outside the student's learning capacity. To address this limitation, we propose Gen-SSD (Generation-time Self-Selection Distillation), a student-in-the-loop framework that performs generation-time selection. Instead of passively consuming complete trajectories, the student evaluates candidate continuations during the teacher's sampling process, guiding the expansion of only learnable reasoning paths and enabling early pruning of unhelpful branches. Experiments on mathematical reasoning benchmarks demonstrate that Gen-SSD consistently outperforms standard knowledge distillation and recent baselines, with improvements of around 5.9 points over Standard KD and up to 4.7 points over other baselines. Further analysis shows that Gen-SSD produces more stable and learnable reasoning trajectories, highlighting the importance of incorporating supervision during generation for effective distillation.

Meta-TTRL: A Metacognitive Framework for Self-Improving Test-Time Reinforcement Learning in Unified Multimodal Models

Existing test-time scaling (TTS) methods for unified multimodal models (UMMs) in text-to-image (T2I) generation primarily rely on search or sampling strategies that produce only instance-level improvements, limiting the ability to learn from prior inferences and accumulate knowledge across similar prompts. To overcome these limitations, we propose Meta-TTRL, a metacognitive test-time reinforcement learning framework. Meta-TTRL performs test-time parameter optimization guided by model-intrinsic monitoring signals derived from the meta-knowledge of UMMs, achieving self-improvement and capability-level improvement at test time. Extensive experiments demonstrate that Meta-TTRL generalizes well across three representative UMMs, including Janus-Pro-7B, BAGEL, and Qwen-Image, achieving significant gains on compositional reasoning tasks and multiple T2I benchmarks with limited data. We provide the first comprehensive analysis to investigate the potential of test-time reinforcement learning (TTRL) for T2I generation in UMMs. Our analysis further reveals a key insight underlying effective TTRL: metacognitive synergy, where monitoring signals align with the model's optimization regime to enable self-improvement.

FlowPrefill: Decoupling Preemption from Prefill Scheduling Granularity to Mitigate Head-of-Line Blocking in LLM Serving

The growing demand for large language models (LLMs) requires serving systems to handle many concurrent requests with diverse service level objectives (SLOs). This exacerbates head-of-line (HoL) blocking during the compute-intensive prefill phase, where long-running requests monopolize resources and delay higher-priority ones, leading to widespread time-to-first-token (TTFT) SLO violations. While chunked prefill enables interruptibility, it introduces an inherent trade-off between responsiveness and throughput: reducing chunk size improves response latency but degrades computational efficiency, whereas increasing chunk size maximizes throughput but exacerbates blocking. This necessitates an adaptive preemption mechanism. However, dynamically balancing execution granularity against scheduling overheads remains a key challenge. In this paper, we propose FlowPrefill, a TTFT-goodput-optimized serving system that resolves this conflict by decoupling preemption granularity from scheduling frequency. To achieve adaptive prefill scheduling, FlowPrefill introduces two key innovations: 1) Operator-Level Preemption, which leverages operator boundaries to enable fine-grained execution interruption without the efficiency loss associated with fixed small chunking; and 2) Event-Driven Scheduling, which triggers scheduling decisions only upon request arrival or completion events, thereby supporting efficient preemption responsiveness while minimizing control-plane overhead. Evaluation on real-world production traces shows that FlowPrefill improves maximum goodput by up to 5.6$\times$ compared to state-of-the-art systems while satisfying heterogeneous SLOs.

DR-LoRA: Dynamic Rank LoRA for Mixture-of-Experts Adaptation

Mixture-of-Experts (MoE) has become a prominent paradigm for scaling Large Language Models (LLMs). Parameter-efficient fine-tuning (PEFT), such as LoRA, is widely adopted to adapt pretrained MoE LLMs to downstream tasks. However, existing approaches assign identical LoRA ranks to all experts, overlooking the intrinsic functional specialization within MoE LLMs. This uniform allocation leads to resource mismatch, task-relevant experts are under-provisioned while less relevant ones receive redundant parameters. We propose a Dynamic Rank LoRA framework named DR-LoRA, which dynamically grows expert LoRA ranks during fine-tuning based on task-specific demands. DR-LoRA employs an Expert Saliency Scoring mechanism that integrates expert routing frequency and LoRA rank importance to quantify each expert's demand for additional capacity. Experts with higher saliency scores are prioritized for rank expansion, enabling the automatic formation of a heterogeneous rank distribution tailored to the target task. Experiments on multiple benchmarks demonstrate that DR-LoRA consistently outperforms standard LoRA and static allocation strategies under the same parameter budget, achieving superior task performance with more efficient parameter utilization.

d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Models

Reliable reinforcement learning (RL) for diffusion large language models (dLLMs) requires both accurate advantage estimation and precise estimation of prediction probabilities. Existing RL methods for dLLMs fall short in both aspects: they rely on coarse or unverifiable reward signals, and they estimate prediction probabilities without accounting for the bias relative to the true, unbiased expected prediction probability that properly integrates over all possible decoding orders. To mitigate these issues, we propose \emph{d}-TreeRPO, a reliable RL framework for dLLMs that leverages tree-structured rollouts and bottom-up advantage computation based on verifiable outcome rewards to provide fine-grained and verifiable step-wise reward signals. When estimating the conditional transition probability from a parent node to a child node, we theoretically analyze the estimation error between the unbiased expected prediction probability and the estimate obtained via a single forward pass, and find that higher prediction confidence leads to lower estimation error. Guided by this analysis, we introduce a time-scheduled self-distillation loss during training that enhances prediction confidence in later training stages, thereby enabling more accurate probability estimation and improved convergence. Experiments show that \emph{d}-TreeRPO outperforms existing baselines and achieves significant gains on multiple reasoning benchmarks, including +86.2 on Sudoku, +51.6 on Countdown, +4.5 on GSM8K, and +5.3 on Math500. Ablation studies and computational cost analyses further demonstrate the effectiveness and practicality of our design choices.

LiRA: Linguistic Robust Anchoring for Cross-lingual Large Language Models

As large language models (LLMs) rapidly advance, performance on high-resource languages (e.g., English, Chinese) is nearing saturation, yet remains substantially lower for low-resource languages (e.g., Urdu, Thai) due to limited training data, machine-translation noise, and unstable cross-lingual alignment. We introduce LiRA (Linguistic Robust Anchoring for Large Language Models), a training framework that robustly improves cross-lingual representations under low-resource conditions while jointly strengthening retrieval and reasoning. LiRA comprises two modules: (i) Arca (Anchored Representation Composition Architecture), which anchors low-resource languages to an English semantic space via anchor-based alignment and multi-agent collaborative encoding, preserving geometric stability in a shared embedding space; and (ii) LaSR (Language-coupled Semantic Reasoner), which adds a language-aware lightweight reasoning head with consistency regularization on top of Arca's multilingual representations, unifying the training objective to enhance cross-lingual understanding, retrieval, and reasoning robustness. We further construct and release a multilingual product retrieval dataset covering five Southeast Asian and two South Asian languages. Experiments across low-resource benchmarks (cross-lingual retrieval, semantic similarity, and reasoning) show consistent gains and robustness under few-shot and noise-amplified settings; ablations validate the contribution of both Arca and LaSR. Code will be released on GitHub and the dataset on Hugging Face.

REMA: A Unified Reasoning Manifold Framework for Interpreting Large Language Model

Understanding how Large Language Models (LLMs) perform complex reasoning and their failure mechanisms is a challenge in interpretability research. To provide a measurable geometric analysis perspective, we define the concept of the Reasoning Manifold, a latent low-dimensional geometric structure formed by the internal representations corresponding to all correctly reasoned generations. This structure can be conceptualized as the embodiment of the effective thinking paths that the model has learned to successfully solve a given task. Based on this concept, we build REMA, a framework that explains the origins of failures by quantitatively comparing the spatial relationships of internal model representations corresponding to both erroneous and correct reasoning samples. Specifically, REMA first quantifies the geometric deviation of each erroneous representation by calculating its k-nearest neighbors distance to the approximated manifold formed by correct representations, thereby providing a unified failure signal. It then localizes the divergence points where these deviations first become significant by tracking this deviation metric across the model's layers and comparing it against a baseline of internal fluctuations from correct representations, thus identifying where the reasoning chain begins to go off-track. Our extensive experiments on diverse language and multimodal models and tasks demonstrate the low-dimensional nature of the reasoning manifold and the high separability between erroneous and correct reasoning representations. The results also validate the effectiveness of the REMA framework in analyzing the origins of reasoning failures. This research connects abstract reasoning failures to measurable geometric deviations in representations, providing new avenues for in-depth understanding and diagnosis of the internal computational processes of black-box models.