Runsong Zhao

Xin Liu, Runsong Zhao, Xinyu Liu et al.

Extended chain-of-thought (CoT) traces improve LLM reasoning but incur substantial computational and memory costs. While existing CoT compression methods mitigate this by condensing thought steps into compact representations via memory tokens and retaining only these representations at inference time, the loss of fine-grained information makes subsequent steps more error-prone. To alleviate this, we propose \textbf{HybridThinker}, where in addition to preserved these representations, thought steps are also temporarily retained to provide fine-grained details. However, we observe that naively keeping thought steps accessible to subsequent steps \emph{during training} lets the model bypass memory tokens by retrieving information directly from these steps, leaving the model's ability to compress and retrieve information through memory tokens insufficiently trained. We therefore introduce a hybrid training scheme, in which only some thought steps are directly accessible through attention to subsequent steps, while the other thought steps are masked, forcing the model to use memory tokens for compression and retrieval. Across 4 reasoning benchmarks, HybridThinker matches the uncompressed baseline, advancing the state of the art in CoT compression by 5.8 points on average accuracy with similar inference time. Ablation studies confirm that both temporary thought-step retention and the hybrid training scheme contribute to these gains.

Runsong Zhao, Shilei Liu, Jiwei Tang et al.

The quadratic complexity and indefinitely growing key-value (KV) cache of standard Transformers pose a major barrier to long-context processing. To overcome this, we introduce the Collaborative Memory Transformer (CoMeT), a novel architecture that enables LLMs to handle arbitrarily long sequences with constant memory usage and linear time complexity. Designed as an efficient, plug-in module, CoMeT can be integrated into pre-trained models with only minimal fine-tuning. It operates on sequential data chunks, using a dual-memory system to manage context: a temporary memory on a FIFO queue for recent events, and a global memory with a gated update rule for long-range dependencies. These memories then act as a dynamic soft prompt for the next chunk. To enable efficient fine-tuning on extremely long contexts, we introduce a novel layer-level pipeline parallelism strategy. The effectiveness of our approach is remarkable: a model equipped with CoMeT and fine-tuned on 32k contexts can accurately retrieve a passkey from any position within a 1M token sequence. On the SCROLLS benchmark, CoMeT surpasses other efficient methods and achieves performance comparable to a full-attention baseline on summarization tasks. Its practical effectiveness is further validated on real-world agent and user behavior QA tasks. The code is available at: https://anonymous.4open.science/r/comet-B00B/

Jiwei Tang, Shilei Liu, Zhicheng Zhang et al.

Large Language Models (LLMs) demonstrate exceptional capability across diverse tasks. However, their deployment in long-context scenarios is hindered by two challenges: computational inefficiency and redundant information. We propose RAM (Read As HuMan), a context compression framework that adopts an adaptive hybrid reading strategy, to address these challenges. Inspired by human reading behavior (i.e., close reading important content while skimming less relevant content), RAM partitions the context into segments and encodes them with the input query in parallel. High-relevance segments are fully retained (close reading), while low-relevance ones are query-guided compressed into compact summary vectors (skimming). Both explicit textual segments and implicit summary vectors are concatenated and fed into decoder to achieve both superior performance and natural language format interpretability. To refine the decision boundary between close reading and skimming, we further introduce a contrastive learning objective based on positive and negative query-segment pairs. Experiments demonstrate that RAM outperforms existing baselines on multiple question answering and summarization benchmarks across two backbones, while delivering up to a 12x end-to-end speedup on long inputs (average length 16K; maximum length 32K).

Runsong Zhao, Xin Liu, Xinyu Liu et al.

Using special tokens (e.g., gist, memory, or compressed tokens) to compress context information is a common practice for large language models (LLMs). However, existing approaches often neglect that position encodings inherently induce local inductive biases in models, causing the compression process to ignore holistic contextual dependencies. We propose \textbf{Enhanced Position Layout (EPL)}, a simple yet effective method that improves the context compression capability of LLMs by only adjusting position IDs, the numerical identifiers that specify token positions. EPL minimizes the distance between context tokens and their corresponding special tokens and at the same time maintains the sequence order in position IDs between context tokens, special tokens, and the subsequent tokens. Integrating EPL into our best performing context compression model results in a 1.9 ROUGE-1 F1 improvement on out-of-domain question answering datasets on average. When extended to multimodal scenarios, EPL leads to an average accuracy gain of 2.6 points for vision compression LLMs.

Runsong Zhao, Shilei Liu, Jiwei Tang et al.

While context compression can mitigate the growing inference costs of Large Language Models (LLMs) by shortening contexts, existing methods that specify a target compression ratio or length suffer from unpredictable performance degradation, hindering their reliable deployment. We introduce a paradigm shift to Performance-oriented Context Compression (PoC), where developers specify an acceptable performance floor instead of a compression ratio. PoC employs a lightweight performance predictor to automatically find the most aggressive compression ratio that satisfies this constraint before steering an off-the-shelf compressor. We design and compare two predictor variants: a simple context-agnostic predictor and a more sophisticated context-aware one that considers the input's inherent compressibility. On both question-answering and summarization benchmarks, the context-aware predictor consistently achieves lower performance prediction error than the context-agnostic predictor, while the resulting context-aware PoC attains a superior overall performance. Our work paves the way for a more reliable, efficient, and performance-aware deployment of context compression for LLMs.

Xinyu Liu, Xin Liu, Bo Jin et al.

While Chain-of-thought (CoT) reasoning enables LLMs to solve challenging reasoning problems, as KV cache grows linearly with the number of generated tokens, CoT reasoning faces scaling issues in terms of speed and memory usage. In this work, we propose MemoSight (Memory-Foresight-based reasoning), a unified framework that integrates both context compression and multi-token prediction to mitigate the efficiency issues while maintaining CoT reasoning performance. Our framework adopts the same minimalist design for both context compression and multi-token prediction via special tokens and their corresponding position layout tailored to each token type. Comprehensive experiments on four reasoning benchmarks demonstrate that MemoSight reduces the KV cache footprint by up to 66% and accelerates inference by 1.56x, while outperforming existing CoT compression methods.

Xinyu Liu, Kechen Jiao, Chunyang Xiao et al.

The convergence of reinforcement learning and imitation learning has positioned Reverse KL (RKL) as a promising paradigm for on-policy LLM distillation, aiming to unify exploration with teacher supervision. However, we identify a critical limitation: when the student and teacher distributions diverge significantly, standard RKL often fails to yield meaningful improvement due to uninformative negative feedback. To address this inefficiency, we propose Teacher-Guided Policy Optimization (TGPO), an on-policy algorithm that incorporates dense directional guidance by leveraging teacher predictions conditioned on the student's rollout. Because TGPO remains on-policy, the algorithm integrates seamlessly with existing RLVR frameworks without requiring additional data annotation. Experiments on complex reasoning benchmarks demonstrate that TGPO significantly outperforms standard baselines and is robust to different teachers.

Jianjin Wang, Runsong Zhao, Xiaoqian Liu et al.

Current direct speech-to-speech translation methods predominantly employ speech tokens as intermediate representations. However, a single speech token is not dense in semantics, so we generally need multiple tokens to express a complete semantic unit. To address this limitation, we introduce multi-token prediction (MTP) loss into speech-to-unit translation (S2UT) models, enabling models to predict multiple subsequent tokens at each position, thereby capturing more complete semantics and enhancing information density per position. Initial MTP implementations apply the loss at the final layer, which improves output representation but initiates information enrichment too late. We hypothesize that advancing the information enrichment process to intermediate layers can achieve earlier and more effective enhancement of hidden representation. Consequently, we propose MTP-S2UT loss, applying MTP loss to hidden representation where CTC loss is computed. Experiments demonstrate that all MTP loss variants consistently improve the quality of S2UT translation, with MTP-S2UT achieving the best performance.

Xin Liu, Runsong Zhao, Pengcheng Huang et al.

Context compression presents a promising approach for accelerating large language model (LLM) inference by compressing long contexts into compact representations. Current context compression methods predominantly rely on autoencoding tasks to train context-agnostic compression tokens to compress contextual semantics. While autoencoding tasks enable compression tokens to acquire compression capabilities, compression via autoencoding tasks creates a fundamental mismatch: the models are optimized for reconstruction that diverge from actual downstream tasks, thereby weakening the features more beneficial for real-world usage. We propose Semantic-Anchor Compression (SAC), a novel method that shifts from autoencoding task based compression to an architecture that is equipped with this compression capability \textit{a priori}. Instead of training models to compress contexts through autoencoding tasks, SAC directly selects so-called anchor tokens from the original context and aggregates contextual information into their key-value (KV) representations. By deriving representations directly from the contextual tokens, SAC eliminates the need for autoencoding training. To ensure compression performance while directly leveraging anchor tokens, SAC incorporates two key designs: (1) anchor embeddings that enable the compressor to identify critical tokens, and (2) bidirectional attention modification that allows anchor tokens to capture information from the entire context. Experimental results demonstrate that SAC consistently outperforms existing context compression methods across various compression ratios. On out-of-distribution evaluation using MRQA, SAC achieves 1 EM improvement at 5x compression over strong baselines, with increasing advantages at higher compression ratios.