Existing benchmarks evaluating biases in large language models (LLMs) primarily rely on explicit cues, declaring protected attributes like religion, race, gender by name. However, real-world interactions often contain implicit biases, inferred subtly through names, cultural cues, or traits. This critical oversight creates a significant blind spot in fairness evaluation. We introduce ImplicitBBQ, a benchmark extending the Bias Benchmark for QA (BBQ) with implicitly cued protected attributes across 6 categories. Our evaluation of GPT-4o on ImplicitBBQ illustrates troubling performance disparity from explicit BBQ prompts, with accuracy declining up to 7% in the "sexual orientation" subcategory and consistent decline located across most other categories. This indicates that current LLMs contain implicit biases undetected by explicit benchmarks. ImplicitBBQ offers a crucial tool for nuanced fairness evaluation in NLP.

Abstract--As Large Language Models (LLMs) scale in size and context length, the memory requirements of the key-value (KV) cache have emerged as a major bottleneck during autoregressive decoding. The KV cache grows with sequence length and embedding dimension, often exceeding the memory footprint of the model itself and limiting achievable batch sizes and context windows. T o address this challenge, we present KV-CAR, a unified and architecture-agnostic framework that significantly reduces KV-cache storage while maintaining model fidelity. KV-CAR combines two complementary techniques. First, a lightweight autoencoder learns compact representations of key and value tensors along the embedding dimension, compressing them before they are stored in the KV cache and restoring them upon retrieval. Second, a similarity-driven reuse mechanism identifies opportunities to reuse KV tensors of specific attention heads across adjacent layers. T ogether, these methods reduce the dimensional and structural redundancy in KV tensors without requiring changes to the transformer architecture. Evaluations on GPT -2 and TinyLLaMA models across Wikitext, C4, PIQA, and Winogrande datasets demonstrate that KV-CAR achieves up to 47.85% KV-cache memory reduction with minimal impact on perplexity and zero-shot accuracy. System-level measurements on an NVIDIA A40 GPU show that the reduced KV footprint directly translates into longer sequence lengths and larger batch sizes during inference. Large Language Models (LLMs) have achieved remarkable performance across a wide range of natural language and multimodal tasks due to their ability to capture long-range dependencies and generate contextually rich outputs.

Large Language Model (LLM) agents are emerging to transform daily life. However, existing LLM agents primarily follow a reactive paradigm, relying on explicit user instructions to initiate services, which increases both physical and cognitive workload. In this paper, we propose ProAgent, the first end-to-end proactive agent system that harnesses massive sensory contexts and LLM reasoning to deliver proactive assistance. ProAgent first employs a proactive-oriented context extraction approach with on-demand tiered perception to continuously sense the environment and derive hierarchical contexts that incorporate both sensory and persona cues. ProAgent then adopts a context-aware proactive reasoner to map these contexts to user needs and tool calls, providing proactive assistance. We implement ProAgent on Augmented Reality (AR) glasses with an edge server and extensively evaluate it on a real-world testbed, a public dataset, and through a user study. Results show that ProAgent achieves up to 33.4% higher proactive prediction accuracy, 16.8% higher tool-calling F1 score, and notable improvements in user satisfaction over state-of-the-art baselines, marking a significant step toward proactive assistants. A video demonstration of ProAgent is available at https://youtu.be/pRXZuzvrcVs.

This study addresses the issues of privacy protection and efficiency in instruction fine-tuning of large-scale language models by proposing a parameter-efficient method that integrates differential privacy noise allocation with gradient clipping in a collaborative optimization framework. The method keeps the backbone model frozen and updates parameters through a low-dimensional projection subspace, while introducing clipping and adaptive noise allocation during gradient computation. This design reduces privacy budget consumption and ensures training stability and robustness. The unified framework combines gradient constraints, noise allocation, and parameter projection, effectively mitigating performance fluctuations and privacy risks in multi-task instruction scenarios. Experiments are conducted across hyperparameter, environment, and data sensitivity dimensions. Results show that the method outperforms baseline models in accuracy, privacy budget, and parameter efficiency, and maintains stable performance under diverse and uncertain data conditions. The findings enrich the theoretical integration of differential privacy and parameter-efficient fine-tuning and demonstrate its practical adaptability in instruction tasks, providing a feasible solution for secure training in complex instruction environments.

As large language models become components of larger agentic systems, evaluation reliability becomes critical: unreliable sub-agents introduce brittleness into downstream system behavior. Yet current evaluation practice, reporting a single accuracy number from a single run, obscures the variance underlying these results, making it impossible to distinguish genuine capability improvements from lucky sampling. We propose adopting Intraclass Correlation Coefficient (ICC), a metric from measurement science, to characterize this variance. ICC decomposes observed variance into between-query variance (task difficulty) and within-query variance (agent inconsistency), highlighting whether reported results reflect true capability or measurement noise. We evaluated on GAIA (Levels 1-3, measuring agentic capabilities across varying reasoning complexity) and FRAMES (measuring retrieval and factuality across multiple documents). We found that ICC varies dramatically with task structure, with reasoning and retrieval tasks (FRAMES) exhibit ICC=0.4955-0.7118 across models, and agentic tasks (GAIA) exhibiting ICC=0.304-0.774 across models. For sub-agent replacement decisions in agentic systems, accuracy improvements are only trustworthy if ICC also improves. We demonstrate that ICC converges by n=8-16 trials for structured tasks and n>=32 for complex reasoning, enabling practitioners to set evidence-based resampling budgets. We recommend reporting accuracy alongside ICC and within-query variance as standard practice, and propose updated Evaluation Cards capturing these metrics. By making evaluation stability visible, we aim to transform agentic benchmarking from opaque leaderboard competition to trustworthy experimental science. Our code is open-sourced at https://github.com/youdotcom-oss/stochastic-agent-evals.

Preference alignment has enabled large language models (LLMs) to better reflect human expectations, but current methods mostly optimize for population-level preferences, overlooking individual users. Personalization is essential, yet early approaches-such as prompt customization or fine-tuning-struggle to reason over implicit preferences, limiting real-world effectiveness. Recent "think-then-generate" methods address this by reasoning before response generation. However, they face challenges in long-form generation: their static one-shot reasoning must capture all relevant information for the full response generation, making learning difficult and limiting adaptability to evolving content. To address this issue, we propose FlyThinker, an efficient "think-while-generating" framework for personalized long-form generation. FlyThinker employs a separate reasoning model that generates latent token-level reasoning in parallel, which is fused into the generation model to dynamically guide response generation. This design enables reasoning and generation to run concurrently, ensuring inference efficiency. In addition, the reasoning model is designed to depend only on previous responses rather than its own prior outputs, which preserves training parallelism across different positions-allowing all reasoning tokens for training data to be produced in a single forward pass like standard LLM training, ensuring training efficiency. Extensive experiments on real-world benchmarks demonstrate that FlyThinker achieves better personalized generation while keeping training and inference efficiency.

Personalization is one of the next milestones in advancing AI capability and alignment. We introduce PersonaMem-v2, the state-of-the-art dataset for LLM personalization that simulates 1,000 realistic user-chatbot interactions on 300+ scenarios, 20,000+ user preferences, and 128k-token context windows, where most user preferences are implicitly revealed to reflect real-world interactions. Using this data, we investigate how reinforcement fine-tuning enables a model to improve its long-context reasoning capabilities for user understanding and personalization. We also develop a framework for training an agentic memory system, which maintains a single, human-readable memory that grows with each user over time. In our experiments, frontier LLMs still struggle with implicit personalization, achieving only 37-48% accuracy. While they support long context windows, reasoning remains the bottleneck for implicit personalization tasks. Using reinforcement fine-tuning, we successfully train Qwen3-4B to outperforms GPT-5, reaching 53% accuracy in implicit personalization. Moreover, our agentic memory framework achieves state-of-the-art 55% accuracy while using 16x fewer input tokens, relying on a 2k-token memory instead of full 32k conversation histories. These results underscore the impact of our dataset and demonstrate agentic memory as a scalable path toward real-world personalized intelligence.

We present a mechanistic interpretability study of GPT-2 that causally examines how sentiment information is processed across its transformer layers. Using systematic activation patching across all 12 layers, we test the hypothesized two-stage sentiment architecture comprising early lexical detection and mid-layer contextual integration. Our experiments confirm that early layers (0-3) act as lexical sentiment detectors, encoding stable, position specific polarity signals that are largely independent of context. However, all three contextual integration hypotheses: Middle Layer Concentration, Phenomenon Specificity, and Distributed Processing are falsified. Instead of mid-layer specialization, we find that contextual phenomena such as negation, sarcasm, domain shifts etc. are integrated primarily in late layers (8-11) through a unified, non-modular mechanism. These experimental findings provide causal evidence that GPT-2's sentiment computation differs from the predicted hierarchical pattern, highlighting the need for further empirical characterization of contextual integration in large language models.

Instruction tuning is one of the key steps required for adapting large language models (LLMs) to a broad spectrum of downstream applications. However, this procedure is difficult because real-world datasets are rarely homogeneous; they consist of a mixture of diverse information, causing gradient interference, where conflicting gradients pull the model in opposing directions, degrading performance. A common strategy to mitigate this issue is to group data based on semantic or embedding similarity. However, this fails to capture how data influences model parameters during learning. While recent works have attempted to cluster gradients directly, they randomly project gradients into lower dimensions to manage memory, which leads to accuracy loss. Moreover, these methods rely on expert ensembles which necessitates multiple inference passes and expensive on-the-fly gradient computations during inference. To address these limitations, we propose GradientSpace, a framework that clusters samples directly in full-dimensional gradient space. We introduce an online SVD-based algorithm that operates on LoRA gradients to identify latent skills without the infeasible cost of storing all sample gradients. Each cluster is used to train a specialized LoRA expert along with a lightweight router trained to select the best expert during inference. We show that routing to a single, appropriate expert outperforms expert ensembles used in prior work, while significantly reducing inference latency. Our experiments across mathematical reasoning, code generation, finance, and creative writing tasks demonstrate that GradientSpace leads to coherent expert specialization and consistent accuracy gains over state-of-the-art clustering methods and finetuning techniques.

Large language models (LLMs) face critical safety challenges, as they can be manipulated to generate harmful content through adversarial prompts and jailbreak attacks. Many defenses are typically either black-box guardrails that filter outputs, or internals-based methods that steer hidden activations by operationalizing safety as a single latent feature or dimension. While effective for simple concepts, this assumption is limiting, as recent evidence shows that abstract concepts such as refusal and temporality are distributed across multiple features rather than isolated in one. To address this limitation, we introduce Graph-Regularized Sparse Autoencoders (GSAEs), which extends SAEs with a Laplacian smoothness penalty on the neuron co-activation graph. Unlike standard SAEs that assign each concept to a single latent feature, GSAEs recover smooth, distributed safety representations as coherent patterns spanning multiple features. We empirically demonstrate that GSAE enables effective runtime safety steering, assembling features into a weighted set of safety-relevant directions and controlling them with a two-stage gating mechanism that activates interventions only when harmful prompts or continuations are detected during generation. This approach enforces refusals adaptively while preserving utility on benign queries. Across safety and QA benchmarks, GSAE steering achieves an average 82% selective refusal rate, substantially outperforming standard SAE steering (42%), while maintaining strong task accuracy (70% on TriviaQA, 65% on TruthfulQA, 74% on GSM8K). Robustness experiments further show generalization across LLaMA-3, Mistral, Qwen, and Phi families and resilience against jailbreak attacks (GCG, AutoDAN), consistently maintaining >= 90% refusal of harmful content.