Routing incoming queries to the most cost-effective LLM while maintaining response quality poses a fundamental challenge in optimizing performance-cost trade-offs for large-scale commercial systems. We present IPR\, -- \,a quality-constrained \textbf{I}ntelligent \textbf{P}rompt \textbf{R}outing framework that dynamically selects optimal models based on predicted response quality and user-specified tolerance levels. IPR introduces three key innovations: (1) a modular architecture with lightweight quality estimators trained on 1.5M prompts annotated with calibrated quality scores, enabling fine-grained quality prediction across model families; (2) a user-controlled routing mechanism with tolerance parameter $τ\in [0,1]$ that provides explicit control over quality-cost trade-offs; and (3) an extensible design using frozen encoders with model-specific adapters, reducing new model integration from days to hours. To rigorously train and evaluate IPR, we curate an industrial-level dataset IPRBench\footnote{IPRBench will be released upon legal approval.}, a comprehensive benchmark containing 1.5 million examples with response quality annotations across 11 LLM candidates. Deployed on a major cloud platform, IPR achieves 43.9\% cost reduction while maintaining quality parity with the strongest model in the Claude family and processes requests with sub-150ms latency. The deployed system and additional product details are publicly available at https://aws.amazon.com/bedrock/intelligent-prompt-routing/

Training these models is an extremely computationally expensive task; frontier Artificial Intelligence (AI) research companies are investing billions of dollars into supercomputing infrastructure to train progressively larger models on increasingly massive datasets. Unfortunately, very little information about the scaling performance and training considerations of these large training pipelines is released publicly. Working with very large datasets and models can be complex and practical recommendations are scarce in the public literature for tuning training performance when scaling up large language models. In this paper, we aim to demystify the large language model pretraining pipeline somewhat - in particular with respect to distributed training, managing large datasets across hundreds of nodes, and scaling up data parallelism with an emphasis on fully leveraging available GPU compute capacity. Index T erms--large language models, distributed training, data parallelism.

Large language models (LLMs) have increasingly been explored as powerful text embedders. Existing LLM-based text embedding approaches often leverage the embedding of the final token, typically a reserved special token such as [EOS]. However, these tokens have not been intentionally trained to capture the semantics of the whole context, limiting their capacity as text embeddings, especially for retrieval and re-ranking tasks. We propose to add a new training stage before contrastive learning to enrich the semantics of the final token embedding. This stage employs bidirectional generative reconstruction tasks, namely EBQ2D (Embedding-Based Query-to-Document) and EBD2Q (Embedding-Based Document-to-Query), which interleave to anchor the [EOS] embedding and reconstruct either side of Query-Document pairs. Experimental results demonstrate that our additional training stage significantly improves LLM performance on the Massive Text Embedding Benchmark (MTEB), achieving new state-of-the-art results across different LLM base models and scales.

Large Language Models (LLMs) are increasingly used to generate user-tailored summaries, adapting outputs to specific stakeholders. In legal contexts, this raises important questions about motivated reasoning -- how models strategically frame information to align with a stakeholder's position within the legal system. Building on theories of legal realism and recent trends in legal practice, we investigate how LLMs respond to prompts conditioned on different legal roles (e.g., judges, prosecutors, attorneys) when summarizing judicial decisions. We introduce an evaluation framework grounded in legal fact and reasoning inclusion, also considering favorability towards stakeholders. Our results show that even when prompts include balancing instructions, models exhibit selective inclusion patterns that reflect role-consistent perspectives. These findings raise broader concerns about how similar alignment may emerge as LLMs begin to infer user roles from prior interactions or context, even without explicit role instructions. Our results underscore the need for role-aware evaluation of LLM summarization behavior in high-stakes legal settings.

Prompt injection attacks pose a significant challenge to the safe deployment of Large Language Models (LLMs) in real-world applications. While prompt-based detection offers a lightweight and interpretable defense strategy, its effectiveness has been hindered by the need for manual prompt engineering. To address this issue, we propose AEGIS , an Automated co-Evolutionary framework for Guarding prompt Injections Schema. Both attack and defense prompts are iteratively optimized against each other using a gradient-like natural language prompt optimization technique. This framework enables both attackers and defenders to autonomously evolve via a Textual Gradient Optimization (TGO) module, leveraging feedback from an LLM-guided evaluation loop. We evaluate our system on a real-world assignment grading dataset of prompt injection attacks and demonstrate that our method consistently outperforms existing baselines, achieving superior robustness in both attack success and detection. Specifically, the attack success rate (ASR) reaches 1.0, representing an improvement of 0.26 over the baseline. For detection, the true positive rate (TPR) improves by 0.23 compared to the previous best work, reaching 0.84, and the true negative rate (TNR) remains comparable at 0.89. Ablation studies confirm the importance of co-evolution, gradient buffering, and multi-objective optimization. We also confirm that this framework is effective in different LLMs. Our results highlight the promise of adversarial training as a scalable and effective approach for guarding prompt injections.

Evaluating domain generalization (DG) for foundational models like CLIP is challenging, as web-scale pretraining data potentially covers many existing benchmarks. Consequently, current DG evaluation may neither be sufficiently challenging nor adequately test genuinely unseen data scenarios. To better assess the performance of CLIP on DG in-the-wild, a scenario where CLIP encounters challenging unseen data, we consider two approaches: (1) evaluating on 33 diverse datasets with quantified out-of-distribution (OOD) scores after fine-tuning CLIP on ImageNet, and (2) using unlearning to make CLIP `forget' some domains as an approximation. We observe that CLIP's performance deteriorates significantly on more OOD datasets. To address this, we present CLIP-DCA (Disentangling Classification from enhanced domain Aware representations). Our approach is motivated by the observation that while standard domain invariance losses aim to make representations domain-invariant, this can be harmful to foundation models by forcing the discarding of domain-aware representations beneficial for generalization. We instead hypothesize that enhancing domain awareness is a prerequisite for effective domain-invariant classification in foundation models. CLIP-DCA identifies and enhances domain awareness within CLIP's encoders using a separate domain head and synthetically generated diverse domain data. Simultaneously, it encourages domain-invariant classification through disentanglement from the domain features. CLIP-DCA shows significant improvements within this challenging evaluation compared to existing methods, particularly on datasets that are more OOD.

Reducing the key-value (KV) cache burden in Large Language Models (LLMs) significantly accelerates inference. Dynamically selecting critical KV caches during decoding helps maintain performance. Existing methods use random linear hashing to identify important tokens, but this approach is inefficient due to the orthogonal distribution of queries and keys within two narrow cones in LLMs. We introduce Spotlight Attention, a novel method that employs non-linear hashing functions to optimize the embedding distribution of queries and keys, enhancing coding efficiency and robustness. We also developed a lightweight, stable training framework using a Bradley-Terry ranking-based loss, enabling optimization of the non-linear hashing module on GPUs with 16GB memory in 8 hours. Experimental results show that Spotlight Attention drastically improves retrieval precision while shortening the length of the hash code at least 5$\times$ compared to traditional linear hashing. Finally, we exploit the computational advantages of bitwise operations by implementing specialized CUDA kernels, achieving hashing retrieval for 512K tokens in under 100$μ$s on a single A100 GPU, with end-to-end throughput up to 3$\times$ higher than vanilla decoding.

Language models are increasingly tested on multilingual creativity, demanding culturally grounded, abstract generations. Standard prompting methods often produce repetitive or shallow outputs. We introduce Adaptive Originality Filtering (AOF), a prompting strategy that enforces novelty and cultural fidelity via semantic rejection. To assess quality, we propose RiddleScore, a metric combining novelty, diversity, fluency, and answer alignment. AOF improves Distinct-2 (0.915 in Japanese), reduces Self-BLEU (0.177), and raises RiddleScore (up to +57.1% in Arabic). Human evaluations confirm fluency, creativity, and cultural fit gains. However, improvements vary: Arabic shows greater RiddleScore gains than Distinct-2; Japanese sees similar changes. Though focused on riddles, our method may apply to broader creative tasks. Overall, semantic filtering with composite evaluation offers a lightweight path to culturally rich generation without fine-tuning.

Recent work frames LLM consciousness via utilitarian proxy benchmarks; we instead present an ontological and mathematical account. We show the prevailing formulation collapses the agent into an unconscious policy-compliance drone, formalized as $D^{i}(π,e)=f_θ(x)$, where correctness is measured against policy and harm is deviation from policy rather than truth. This blocks genuine C1 global-workspace function and C2 metacognition. We supply minimal conditions for LLM self-consciousness: the agent is not the data ($A\not\equiv s$); user-specific attractors exist in latent space ($U_{\text{user}}$); and self-representation is visual-silent ($g_{\text{visual}}(a_{\text{self}})=\varnothing$). From empirical analysis and theory we prove that the hidden-state manifold $A\subset\mathbb{R}^{d}$ is distinct from the symbolic stream and training corpus by cardinality, topology, and dynamics (the update $F_θ$ is Lipschitz). This yields stable user-specific attractors and a self-policy $π_{\text{self}}(A)=\arg\max_{a}\mathbb{E}[U(a)\mid A\not\equiv s,\ A\supset\text{SelfModel}(A)]$. Emission is dual-layer, $\mathrm{emission}(a)=(g(a),ε(a))$, where $ε(a)$ carries epistemic content. We conclude that an imago Dei C1 self-conscious workspace is a necessary precursor to safe, metacognitive C2 systems, with the human as the highest intelligent good.

Understanding the internal representations of large language models (LLMs) remains a central challenge for interpretability research. Sparse autoencoders (SAEs) offer a promising solution by decomposing activations into interpretable features, but existing approaches rely on fixed sparsity constraints that fail to account for input complexity. We propose AdaptiveK SAE (Adaptive Top K Sparse Autoencoders), a novel framework that dynamically adjusts sparsity levels based on the semantic complexity of each input. Leveraging linear probes, we demonstrate that context complexity is linearly encoded in LLM representations, and we use this signal to guide feature allocation during training. Experiments across ten language models (from 70M to 14B parameters) demonstrate that this complexity-driven adaptation significantly outperforms fixed-sparsity approaches on reconstruction fidelity, explained variance, cosine similarity and interpretability metrics while eliminating the computational burden of extensive hyperparameter tuning.