We prove that any Turing machine running on inputs of arbitrary length can be simulated by a constant bit-size transformer, as long as the context window is sufficiently long. This improves previous works, which require scaling up either the model's precision or the number of parameters on longer inputs. Furthermore, we prove that the complexity class SPACE$[s(n)]$ exactly characterizes the expressive power of a constant bit-size transformer with a context window of length $s(n)$. Our approach relies on simulating Post machines, a Turing-complete computational model. Post machines can be modeled as automata equipped with a queue, exhibiting computational behaviors naturally aligned with those of transformers. The behavioral similarity between transformers and Post machines may offer new insights into the mechanisms underlying the reasoning abilities of transformers.

As the cost of pretraining large language models grows, there is continued interest in strategies to improve learning efficiency during this core training stage. Motivated by cognitive development, where humans gradually build knowledge as their brains mature, we propose Curriculum-Guided Layer Scaling (CGLS), a framework for compute-efficient pretraining that synchronizes increasing data difficulty with model growth through progressive layer stacking (i.e., gradually adding layers during training). At the 100M parameter scale, using a curriculum transitioning from synthetic short stories to general web data, CGLS outperforms baseline methods on the question-answering benchmarks PIQA and ARC. Our results show that progressively increasing model depth alongside sample difficulty leads to better generalization and zero-shot performance on various downstream benchmarks. Altogether, our findings demonstrate that CGLS unlocks the potential of progressive stacking, offering a simple yet effective strategy for improving generalization on knowledge-intensive and reasoning tasks. Large language models (LLMs) are typically pretrained in a single, continuous pass, processing all tokens with a uniform amount of computation regardless of their complexity or relevance to downstream tasks of interest. While this approach has shown remarkable success in large-scale models like GPT -4 (OpenAI et al., 2023) and Llama 3 (Dubey et al., 2024), it differs significantly from how humans learn, often leading to models that excel in generating coherent text but struggle with long-context reasoning across varied tasks (Schnabel et al., 2025). Recent works like Phi-3 (Abdin et al., 2024), MiniCPM (Hu et al., 2024), and others (Feng et al., 2024) have explored midtraining, adjusting the training data distribution partway through training by incorporating higher-quality, multilingual, or long-form text. However, this coarse-grained curriculum is applied on fixed model architectures. Inspired by how humans progressively build knowledge alongside their physically growing brains, we explore whether gradually scaling a model in tandem with increasingly complex data can enable more efficient and effective learning.

Fine-tuning is a crucial process for adapting large language models (LLMs) to diverse applications. In certain scenarios, like multi-tenant serving, a large number of LLMs finetuned from the same base model are deployed to meet complex requirements for users. Recent works explore delta-compression approaches to quantize and compress the delta weights between the customized LLM and the corresponding base model. However, they exhibit inadequate performance at high compression ratios due to their empirical nature. In this work, we introduce DeltaMix, an adaptive mixed-precision delta-compression framework designed to minimize quantization error in the singular value decomposition (SVD) space without imposing additional assumptions. DeltaMix provides a theoretical justification for the necessity of mixed-precision compression and presents a practical quantization solution that involves solving a 0/1 linear integer programming problem alongside a reconstruction target correction method. Experimental results across multiple models and benchmarks illustrate that DeltaMix consistently outperforms all baseline methods. Notably, on tasks such as AIME2024 and GQA, DeltaMix exceeds the performance of the best baseline, Delta-CoMe, by 22.3\% and 6.1\% for 7B parameter models, respectively.

Discrete audio tokens are compact representations that aim to preserve perceptual quality, phonetic content, and speaker characteristics while enabling efficient storage and inference, as well as competitive performance across diverse downstream tasks. They provide a practical alternative to continuous features, enabling the integration of speech and audio into modern large language models (LLMs). As interest in token-based audio processing grows, various tokenization methods have emerged, and several surveys have reviewed the latest progress in the field. However, existing studies often focus on specific domains or tasks and lack a unified comparison across various benchmarks. This paper presents a systematic review and benchmark of discrete audio tokenizers, covering three domains: speech, music, and general audio. We propose a taxonomy of tokenization approaches based on encoder-decoder, quantization techniques, training paradigm, streamability, and application domains. We evaluate tokenizers on multiple benchmarks for reconstruction, downstream performance, and acoustic language modeling, and analyze trade-offs through controlled ablation studies. Our findings highlight key limitations, practical considerations, and open challenges, providing insight and guidance for future research in this rapidly evolving area.

LLM agents have begun to appear as personal assistants, customer service bots, and clinical aides. While these applications deliver substantial operational benefits, they also require continuous access to sensitive data, which increases the likelihood of unauthorized disclosures. Moreover, these disclosures go beyond mere explicit disclosure, leaving open avenues for gradual manipulation or sidechannel information leakage. This study proposes an auditing framework for conversational privacy that quantifies an agent's susceptibility to these risks. The proposed Conversational Manipulation for Privacy Leakage (CMPL) framework is designed to stress-test agents that enforce strict privacy directives against an iterative probing strategy. Rather than focusing solely on a single disclosure event or purely explicit leakage, CMPL simulates realistic multi-turn interactions to systematically uncover latent vulnerabilities. Our evaluation on diverse domains, data modalities, and safety configurations demonstrates the auditing framework's ability to reveal privacy risks that are not deterred by existing single-turn defenses, along with an in-depth longitudinal study of the temporal dynamics of leakage, strategies adopted by adaptive adversaries, and the evolution of adversarial beliefs about sensitive targets. In addition to introducing CMPL as a diagnostic tool, the paper delivers (1) an auditing procedure grounded in quantifiable risk metrics and (2) an open benchmark for evaluation of conversational privacy across agent implementations.

Recent work has identified retrieval heads, a subset of attention heads responsible for retrieving salient information in long-context language models (LMs), as measured by their copy-paste behavior in Needlein-a-Haystack tasks. In this paper, we introduce QRHead (Query-Focused Retrieval Head), an improved set of attention heads that enhance retrieval from long context. We identify QRHead by aggregating attention scores with respect to the input query, using a handful of examples from real-world tasks (e.g., long-context QA). We further introduce QRRetriever, an efficient and effective retriever that uses the accumulated attention mass of QRHead as retrieval scores. We use QRRetriever for long-context reasoning by selecting the most relevant parts with the highest retrieval scores. On multi-hop reasoning tasks LongMemEval and CLIPPER, this yields over 10% performance gains over full context and outperforms strong dense retrievers. We also evaluate QRRetriever as a re-ranker on the BEIR benchmark and find that it achieves strong zero-shot performance, outperforming other LLM-based re-rankers such as RankGPT. Further analysis shows that both the query-context attention scoring and task selection are crucial for identifying QRHead with strong downstream utility. Overall, our work contributes a general-purpose retriever and offers interpretability insights into the long-context capabilities of LMs.

Despite the rapid progress of multimodal large language models (MLLMs), they have largely overlooked the importance of visual processing. In a simple yet revealing experiment, we interestingly find that language-only models, when provided with image captions, can achieve comparable or even better performance than MLLMs that consume raw visual inputs. This suggests that current MLLMs may generate accurate visual descriptions but fail to effectively integrate them during reasoning. Motivated by this, we propose a simple visual perturbation framework that enhances perceptual robustness without requiring algorithmic modifications or additional training data. Our approach introduces three targeted perturbations: distractor concatenation, dominance-preserving mixup, and random rotation, that can be easily integrated into existing post-training pipelines including SFT, DPO, and GRPO. Through extensive experiments across multiple datasets, we demonstrate consistent improvements in mathematical reasoning performance, with gains comparable to those achieved through algorithmic changes. Additionally, we achieve competitive performance among open-source 7B RL-tuned models by training Qwen2.5-VL-7B with visual perturbation. Through comprehensive ablation studies, we analyze the effectiveness of different perturbation strategies, revealing that each perturbation type contributes uniquely to different aspects of visual reasoning. Our findings highlight the critical role of visual perturbation in multimodal mathematical reasoning: better reasoning begins with better seeing. Our code is available at https://github.com/YutingLi0606/Vision-Matters.

Designing ligand-binding proteins with precise functions is fundamental to advances in biology and chemistry, yet existing AI approaches are limited by scarce protein-ligand complex data. Meanwhile, abundant text descriptions of protein-ligand interactions remain underutilized. We introduce InstructPro, a family of generative models that design proteins from natural language instructions and ligand formulas. InstructPro produces protein sequences consistent with specified functional descriptions and ligand targets. To enable training and evaluation, we develop InstructProBench, a large-scale dataset of 9.6 million (function description, ligand, protein) triples. We train two model variants: InstructPro-1B and InstructPro-3B, which substantially outperform strong baselines. InstructPro-1B achieves design success rates of 2.46% (seen ligands) and 3.14% (zero-shot), while InstructPro-3B reaches 5.06% and 3.93%, respectively. These results demonstrate the potential of natural language-guided generative modeling to expand protein design capabilities beyond traditional data limitations.

Adversarial threats against LLMs are escalating faster than current defenses can adapt. We expose a critical geometric blind spot in alignment: adversarial prompts exploit latent camouflage, embedding perilously close to the safe representation manifold while encoding unsafe intent thereby evading surface level defenses like Direct Preference Optimization (DPO), which remain blind to the latent geometry. We introduce ALKALI, the first rigorously curated adversarial benchmark and the most comprehensive to date spanning 9,000 prompts across three macro categories, six subtypes, and fifteen attack families. Evaluation of 21 leading LLMs reveals alarmingly high Attack Success Rates (ASRs) across both open and closed source models, exposing an underlying vulnerability we term latent camouflage, a structural blind spot where adversarial completions mimic the latent geometry of safe ones. To mitigate this vulnerability, we introduce GRACE - Geometric Representation Aware Contrastive Enhancement, an alignment framework coupling preference learning with latent space regularization. GRACE enforces two constraints: latent separation between safe and adversarial completions, and adversarial cohesion among unsafe and jailbreak behaviors. These operate over layerwise pooled embeddings guided by a learned attention profile, reshaping internal geometry without modifying the base model, and achieve up to 39% ASR reduction. Moreover, we introduce AVQI, a geometry aware metric that quantifies latent alignment failure via cluster separation and compactness. AVQI reveals when unsafe completions mimic the geometry of safe ones, offering a principled lens into how models internally encode safety. We make the code publicly available at https://anonymous.4open.science/r/alkali-B416/README.md.

C. VLA Post-training Framework in Simulation After improving the computation of relative advantages and deriving the corresponding optimization objective, we integrated these components into a complete online reinforcement learning framework for VLA post-training in simulation. First, our overall framework trains a VLA model for a single task using reinforcement learning across multiple environments initialized with identical states. In this setup, the VLA executes the same task in parallel environments, sampling actions step by step until either one environment completes the task or all environments reach the maximum number of steps. During sampling, we incorporate the multistage reward function designed by the LLM described earlier, where each environment's observations provide the necessary object positions and robot state information required for reward computation. Once the trajectories terminate simultaneously, they all share the same length, which facilitates consistent grouping for subsequent processing. After collecting multiple trajectories, they are organized into a trajectory-level group, where relative advantages are computed within the group according to Eq. (3), yielding trajectory-level relative advantages. Similarly, since all trajectories terminate at the same timestep (ensuring that every step can be grouped), we extract step-level data across trajectories (e.g., rewards and log probabilities of actions), and group together steps at the same timestep to form step-level groups. Step-level relative advantages are then computed using Eq.