Large language models (LLMs) remain vulnerable to jailbreak attacks that elicit structured, credible misinformation. Existing benchmarks primarily track binary refusals, often overlooking the severity and exposure risk of the misinformation payload itself. In this work, we introduce Misinfo-Jailbreak Bench, a comprehensive framework that systematically maps diverse attack strategies to a claim-level scoring pipeline for precise risk quantification. Building on this diagnostic foundation, we propose Selective LoRA Ethical Surgery, a parameter-efficient defense mechanism. Instead of broad parameter updates, our approach identifies safety-relevant subspaces via contrastive representations to confine LoRA edits and employs inference-time gating to preserve benign capabilities. Experimental results on state-of-the-art models (e.g., Qwen2.5, Llama-4) demonstrate that our method reduces high-severity risk by 60-70% while limiting general capability degradation to less than 1%, offering a precise and minimally invasive paradigm for post-hoc alignment.

Parametric retrieval-augmented generation (PRAG) injects retrieved evidence into a large language model (LLM) through passage-specific LoRA adapters, reducing reliance on long in-context prompts. When multiple passages are retrieved for the same query, however, evidence-level fusion becomes a bottleneck: equal-weight merging can amplify weak or conflicting evidence, and translating retrieval signals into fusion weights often requires fragile global tuning. We propose FCPRAG, a fusion-controlled parametric RAG framework that adds a lightweight controller for retrieval-conditioned, sample-level adapter fusion. The controller predicts per-passage fusion scores together with sample-level calibration signals, including a mixing gate and an adaptive temperature, enabling fusion that stays selective under informative retrieval signals and conservative under uncertainty. FCPRAG is trained with merge-aware supervision derived from each adapter’s marginal contribution within a multi-adapter merge, using training data only. We further show that a single dataset-level temperature is suboptimal under heteroscedastic retrieval uncertainty, motivating sample-level adaptation. Experiments on HotpotQA, 2WikiMultiHopQA, PopQA, and ComplexWebQuestions (CWQ) across three LLM backbones show that FCPRAG consistently improves F1 over standard RAG and parametric RAG baselines, with gains of up to 4.65% on 2WikiMultiHopQA and 7.55% on CWQ, while also reducing tuning cost and improving robustness under retrieval perturbations. Code is available at https://github.com/ReinerBRO/FCPRAG.

Multimodal models often align language, vision, and audio in a single final-layer latent space, implicitly assuming that task-relevant evidence emerges at the same semantic depth across modalities. Using layer-wise CKA analysis, we observe that this assumption leads to semantic granularity mismatch: textual cues usually require deeper contextual abstraction, whereas visual and acoustic cues often provide discriminative perceptual evidence in shallow or middle layers. This mismatch can flatten fine-grained modality-private cues and reduce reliability under noisy, imbalanced, or missing inputs. To address this, we proposed Adaptive Hierarchical Representation Alliance (AHRA), a hierarchical shared–private expert framework. AHRA factorizes each modality into shared and private streams across semantic levels, regularizes them with shared alignment and private decorrelation, routes shared information through a cross-modal expert, and enhances task-relevant private tokens with modality-specific experts guided by a sparsity-controlled soft-gating mechanism (foreground exam). A hierarchical co-fusion module then performs intra-level expert coordination and inter-level semantic selection. Experiments on six benchmarks across image-text classification, multimodal intent recognition, and trimodal sentiment analysis show that AHRA consistently improves over strong baselines and remains robust under noisy and missing-modality settings.

Extending large language models (LLMs) to low-resource languages often incurs an “align- ment tax”: improvements in the target lan- guage come at the cost of catastrophic forget- ting in general capabilities. We argue that this trade-off arises from the rigidity of supervised fine-tuning (SFT), which enforces token-level surface imitation on narrow and biased data distributions. To address this limitation, we propose a semantic-space alignment paradigm powered by Group Relative Policy Optimiza- tion (GRPO), where the model is optimized us- ing embedding-level semantic rewards rather than likelihood maximization. This objective encourages meaning preservation through flex- ible realizations, enabling controlled updates that reduce destructive interference with pre- trained knowledge. We evaluate our approach on Tibetan–Chinese machine translation and Ti- betan headline generation. Experiments show that our method acquires low-resource capa- bilities while markedly mitigating alignment tax, preserving general competence more effec- tively than SFT. Despite producing less rigid surface overlap, semantic RL yields higher se- mantic quality and preference in open-ended generation, and few-shot transfer results indi- cate that it learns more transferable and ro- bust representations under limited supervision. Overall, our study demonstrates that reinforce- ment learning with semantic rewards provides a safer and more reliable pathway for inclusive low-resource language expansion.

Reinforcement learning from verifiable rewards (RLVR) drives chain-of-thought reasoning in large language models, yet its binary outcome reward cannot distinguish among correct trajectories. Existing dense reward alternatives, from surface heuristics to process reward models, either ignore the expert solutions already present in training corpora or require expensive offline annotation. We propose Gradient-Aligned Reward (GAR), which operates in the policy’s own gradient space: truncated backpropagation through the output projection layer extracts a compact gradient vector for each rollout, and cosine similarity with an expert-anchor gradient yields a dense, reasoning-aware reward with less than 9% wall-clock overhead. We prove that this cosine admits a multiplicative decomposition into prediction-error and activation-pattern factors, providing a concrete characterization of what the alignment signal measures. On Qwen3-4B and Qwen3-8B, GAR consistently improves over GRPO and other baselines on competition-level math benchmarks and transfers to GPQA Diamond and MMLU-Pro without domain-specific data. Code and data are available at https://github.com/LQgdwind/GAR .

The application of autonomous agents to scientific literature review is often constrained by the reliability of online search infrastructures and linear text generation. To address these limitations, we introduce SciLENS (Scientific Localized Evidence Navigation and Synthesis), an autonomous agent framework operating entirely upon a localized dual-tier data infrastructure for deterministic access to massive academic corpora. To mitigate online dependency, SciLENS utilizes a comprehensive offline toolbox featuring topological graph traversal and structural visualization. This empowers the agent to navigate deep citation networks and organize macro-level trends into structured visual representations, mitigating context exhaustion. To train this system without human annotation, we develop an automated multi-hop data synthesis pipeline validated by cross-model consensus, followed by rubric-based reinforcement learning to align the logical planning of the agent. Evaluations across standard tasks and newly constructed complex synthesis benchmarks demonstrate that SciLENS achieves robust factual precision, outperforming existing open-source baselines and achieving performance comparable to frontier proprietary models. Our source code and data are released at https://anonymous.4open.science/r/SciLENS-48DB .

Criminal judgment prediction requires models to infer statutory articles, charges, and sentencing outcomes from case facts. Unlike standard classification tasks, it involves a structured reasoning process in which statutes should be matched with facts, charges should be justified by statutes, and sentencing outcomes should remain consistent with charges. Existing approaches optimize final labels, and while some have attempted to evaluate reasoning quality, their evaluations are indirect, often relying on LLM-generated rubrics that reflect model-internal preferences rather than the inherent logical structure of legal adjudication. We propose Juris Policy Optimization (JPO), a post-training framework for structured legal reasoning in Chinese criminal judgment prediction. JPO first uses teacher-generated rationales to supervise a standardized four-step reasoning process, and then applies reinforcement learning with a composite reward over legal prediction quality, reasoning structure completeness, and cross-step consistency. JPO further introduces token-level advantage reweighting and adaptive clipping for legally salient reasoning segments. Experiments on multiple open-source language models and three Chinese legal benchmarks show that JPO consistently improves both judgment prediction and reasoning quality over supervised fine-tuning and reinforcement learning baselines.