Fine-tuning Large Language Models (LLMs) on multi-turn reasoning datasets requires N (number of turns) separate forward passes per conversation due to reasoning token visibility constraints, as reasoning tokens for a turn are discarded in subsequent turns. We propose duplicating response tokens along with a custom attention mask to enable single-pass processing of entire conversations. We prove our method produces identical losses to the N-pass approach while reducing time complexity from $O\bigl(N^{3}\bigl)$ to $O\bigl(N^{2}\bigl)$ and maintaining the same memory complexity for a transformer based model. Our approach achieves significant training speedup while preserving accuracy. Our implementation is available online (https://github.com/devrev/One-Pass-to-Reason).

Recent Multimodal Large Language Models (MLLMs) have demonstrated strong performance in visual grounding, establishing themselves as a general interface for various vision-language applications. This progress has driven the development of token pruning methods to mitigate the high computational costs associated with processing numerous visual tokens. However, we observe that pruning significantly weakens the model's grounding ability, leading to incorrect predictions and drastic performance degradation. In Referring Expression Comprehension (REC), for instance, pruning causes the accuracy of LLaVA on the RefCOCO validation set to drop from 56.14% to 15.34%. Our analysis identifies misaligned position IDs after pruning as the primary cause of this degradation, as both the order and value of these IDs are crucial for maintaining performance in grounding tasks. To address this issue, we propose Grounding-Aware Token Pruning (GAP), a simple yet effective adjustment to position IDs that recovers REC accuracy back to 51.42%, which is 90% of the original performance in the without pruning setting, all while requiring no additional training, memory, or computational overhead. Applied to models such as Shikra, MiniGPTv2, and the LLaVA series, our method consistently improves performance across various token pruning strategies.

Large Language Models (LLMs) are primarily designed for batch processing. Existing methods for adapting LLMs to streaming rely either on expensive re-encoding or specialized architectures with limited scalability. This work identifies three key mismatches in adapting batch-oriented LLMs to streaming: (1) input-attention, (2) output-attention, and (3) position-ID mismatches. While it is commonly assumed that the latter two mismatches require frequent re-encoding, our analysis reveals that only the input-attention mismatch significantly impacts performance, indicating re-encoding outputs is largely unnecessary. To better understand this discrepancy with the common assumption, we provide the first comprehensive analysis of the impact of position encoding on LLMs in streaming, showing that preserving relative positions within source and target contexts is more critical than maintaining absolute order. Motivated by the above analysis, we introduce a group position encoding paradigm built on batch architectures to enhance consistency between streaming and batch modes. Extensive experiments on cross-lingual and cross-modal tasks demonstrate that our method outperforms existing approaches. Our method requires no architectural modifications, exhibits strong generalization in both streaming and batch modes. The code is available at repository https://github.com/EIT-NLP/StreamingLLM.

Currently, a prevalent approach for enhancing Vision-Language Models (VLMs) performance is to encode both the high-resolution version and the thumbnail of an image simultaneously. While effective, this method generates a large number of image tokens. When combined with the widely used Rotary Position Embedding (RoPE), its long-term decay property hinders the interaction between high-resolution tokens and thumbnail tokens, as well as between text and image. To address these issues, we propose ID-Align, which alleviates these problems by reordering position IDs. In this method, high-resolution tokens inherit IDs from their corresponding thumbnail token while constraining the overexpansion of positional indices. Our experiments conducted within the LLaVA-Next framework demonstrate that ID-Align achieves significant improvements, including a 6.09% enhancement on MMBench's relation reasoning tasks and notable gains across multiple benchmarks. Our code is available at the following link: https://github.com/zooblastlbz/ID-Align.

Large language models (LLMs) that integrate multiple input roles (e.g., system instructions, user queries, external tool outputs) are increasingly prevalent in practice. Ensuring that the model accurately distinguishes messages from each role -- a concept we call \emph{role separation} -- is crucial for consistent multi-role behavior. Although recent work often targets state-of-the-art prompt injection defenses, it remains unclear whether such methods truly teach LLMs to differentiate roles or merely memorize known triggers. In this paper, we examine \emph{role-separation learning}: the process of teaching LLMs to robustly distinguish system and user tokens. Through a \emph{simple, controlled experimental framework}, we find that fine-tuned models often rely on two proxies for role identification: (1) task type exploitation, and (2) proximity to begin-of-text. Although data augmentation can partially mitigate these shortcuts, it generally leads to iterative patching rather than a deeper fix. To address this, we propose reinforcing \emph{invariant signals} that mark role boundaries by adjusting token-wise cues in the model's input encoding. In particular, manipulating position IDs helps the model learn clearer distinctions and reduces reliance on superficial proxies. By focusing on this mechanism-centered perspective, our work illuminates how LLMs can more reliably maintain consistent multi-role behavior without merely memorizing known prompts or triggers.

Training large language models to predict beyond their training context lengths has drawn much attention in recent years, yet the principles driving such behavior of length generalization remain underexplored. We propose a new theoretical framework to study length generalization for the next-token prediction task, as performed by decoder-only transformers. Conceptually, we show that length generalization occurs as long as each predicted token depends on a small (fixed) number of previous tokens. We formalize such tasks via a notion we call $k$-sparse planted correlation distributions, and show that an idealized model of transformers which generalize attention heads successfully length-generalize on such tasks. As a bonus, our theoretical model justifies certain techniques to modify positional embeddings which have been introduced to improve length generalization, such as position coupling. We support our theoretical results with experiments on synthetic tasks and natural language, which confirm that a key factor driving length generalization is a ``sparse'' dependency structure of each token on the previous ones. Inspired by our theory, we introduce Predictive Position Coupling, which trains the transformer to predict the position IDs used in a positional coupling approach. Predictive Position Coupling thereby allows us to broaden the array of tasks to which position coupling can successfully be applied to achieve length generalization.

Decoding with autoregressive large language models (LLMs) traditionally occurs sequentially, generating one token after another. An emerging line of work explored parallel decoding by identifying and simultaneously generating semantically independent chunks of LLM responses. However, these techniques rely on hand-crafted heuristics tied to syntactic structures like lists and paragraphs, making them rigid and imprecise. We present PASTA, a learning-based system that teaches LLMs to identify semantic independence and express parallel decoding opportunities in their own responses. At its core are PASTA-LANG and its interpreter: PASTA-LANG is an annotation language that enables LLMs to express semantic independence in their own responses; the language interpreter acts on these annotations to orchestrate parallel decoding on-the-fly at inference time. Through a two-stage finetuning process, we train LLMs to generate PASTA-LANG annotations that optimize both response quality and decoding speed. Evaluation on AlpacaEval, an instruction following benchmark, shows that our approach Pareto-dominates existing methods in terms of decoding speed and response quality; our results demonstrate geometric mean speedups ranging from 1.21x to 1.93x with corresponding quality changes of +2.2% to -7.1%, measured by length-controlled win rates against sequential decoding baseline.

Mitigating positional bias of language models (LMs) for listwise inputs is a well-known and important problem (e.g., lost-in-the-middle). While zero-shot order-invariant LMs have been proposed to solve this issue, their success on practical listwise problems has been limited. In this work, as a first contribution, we identify and overcome two limitations to make zero-shot invariant LMs more practical: (1) training and inference distribution mismatch arising from modifying positional ID assignments to enforce invariance, and (2) failure to adapt to a mixture of order-invariant and sensitive inputs in practical listwise problems. To overcome, we propose (1) RoToR, a zero-shot invariant LM for genuinely order-invariant inputs with minimal modifications of positional IDs, and (2) Selective Routing, an adaptive framework that handles both order-invariant and order-sensitive inputs in listwise tasks. On the Lost in the middle (LitM), Knowledge Graph Question Answering (KGQA), and MMLU benchmarks, we show that RoToR with Selective Routing can effectively handle practical listwise input tasks in a zero-shot manner.

Extending context window sizes allows large language models (LLMs) to process longer sequences and handle more complex tasks. Rotary Positional Embedding (RoPE) has become the de facto standard due to its relative positional encoding properties that benefit long-context training. However, we observe that using RoPE with BFloat16 format results in numerical issues, causing it to deviate from its intended relative positional encoding, especially in long-context scenarios. This issue arises from BFloat16's limited precision and accumulates as context length increases, with the first token contributing significantly to this problem. To address this, we develop AnchorAttention, a plug-and-play attention method that alleviates numerical issues caused by BFloat16, improves long-context capabilities, and speeds up training. AnchorAttention reduces unnecessary attention computations, maintains semantic coherence, and boosts computational efficiency by treating the first token as a shared anchor with a consistent position ID, making it visible to all documents within the training context. Experiments on three types of LLMs demonstrate that AnchorAttention significantly improves long-context performance and reduces training time by over 50\% compared to standard full attention mechanisms, while preserving the original LLM's capabilities on general tasks. Our code is available at https://github.com/haonan3/AnchorContext.