arXiv subjects

Piji Li

Publications and source records attributed to Piji Li.

At least 19 recordsLinked to original sources

arXiv2026

Enhancing Knowledge Density via Hierarchical Context Pruning for Retrieval-Augmented Repository-Level Code Completion

Retrieval-Augmented Generation (RAG) has advanced repository-level code completion by retrieving external code snippets, API documentation, or type definitions to expand the model's contextual awareness. However, under the finite token budget of the context window, existing retrieval-augmented methods often saturate the budget with highly similar or even redundant code, which not only wastes tokens but also suppresses critical cross-file dependencies, preventing synergistic interaction between semantically similar reference snippets and cross-file dependency definitions. Therefore, further gains are limited by low knowledge density in the retrieved context rather than by model capacity. We propose SaraCoder, a retrieval-augmented framework that enhances knowledge density via hierarchical context pruning under a fixed context budget. A Hierarchical Feature Optimization module prunes an over-complete candidate pool in four stages. The freed tokens are then allocated to an External-Aware Identifier Disambiguator, which selectively injects only the most necessary cross-file dependency definitions. SaraCoder transforms context construction into a budget-aware optimization problem, balancing reference snippets and cross-file dependencies to build a context that is both knowledge-dense and semantically complete. Experiments on CrossCodeEval and RepoEval-Updated, across multiple backbone code LLMs, show that SaraCoder consistently outperforms strong RAG baselines such as GraphCoder and Repocoder in both code-level and identifier-level metrics, without increasing token usage, and yields synergistic gains when combined with other cross-file methods.

cs.SE↗

arXiv2026

GIFT: Goal-Injected Fine-Tuning for Efficient Manipulation Policy Adaptation

Compared with relying solely on initial observations and language instructions, predicting goal images with generative models as high-level visual guidance can significantly enhance the robustness of Vision-Language-Action (VLA) models. However, most existing foundation models have not systematically incorporated goal image conditioning due to the high computational training cost. To this end, we propose Goal-Injected Fine-Tuning (GIFT), a lightweight and efficient fine-tuning framework that seamlessly integrates generated goal images into multiple representative pretrained VLA models. Our approach introduces goal image features into observations via a zero-initialized convolution which progressively grows parameters from zero and prevents harmful noise from disrupting the pretrained policy during fine-tuning. As training proceeds, goal information is gradually incorporated, enabling efficient goal understanding without disrupting model stability. We further introduce a refined image editing method to generate semantically and visually consistent goal images from initial observations and task instructions. Experiments show that goal-aware VLA models achieve substantial performance gains across tasks: with only a single epoch of fine-tuning, GIFT outperforms the base model by 6.0% and 13.4% on two SIMPLER settings, and by 4.7% on LIBERO, demonstrating both efficiency and effectiveness.

cs.CV↗

arXiv2026

MI-Distillation: Selecting from Model-Interpolated Instruct-Reasoning Data Spectrum for Chain-of-Thought Distillation

Recent advances in large reasoning models (LRMs) have shown strong performance on complex problems through long chain-of-thought (Long CoT) reasoning. However, distilling such trajectories into smaller student models remains challenging: direct Long CoT supervision often provides limited gains and can be less effective than concise Short CoT rationales. In this work, we investigate this phenomenon from a gradient-centric perspective. Our analysis shows that Long CoT induces larger gradient magnitudes and more concentrated update directions than Short CoT, with this effect becoming more pronounced as student model capacity increases. These findings suggest that effective Long CoT distillation requires balancing the reasoning information density of reasoning trajectories with their distributional alignment to the student model. Motivated by this insight, we propose \textbf{M}odel \textbf{I}nterporlation \textbf{Distillation} (\textbf{MI-Distillation}), a framework that constructs a continuous Instruct-Reasoning data spectrum through model interpolation. To select suitable trajectories from this spectrum, we further introduce \textbf{Seq}uential \textbf{L}earnable \textbf{S}urprisal \textbf{S}core (\textbf{SeqLSS}), which favors reasoning paths that are both informative and learnable for the student. Extensive experiments on reasoning benchmarks show that MI-Distillation consistently improves small model CoT distillation over strong Long CoT baselines.

cs.CL↗

arXiv2026

ViSMoE: Visual-Aware Sparse Mixture-of-Experts for Embodied Referring Expression Grounding

Embodied Referring Expression Grounding is the task of enabling an agent to navigate in real environments and to localize a remote object based on natural language instructions. In this scenario, the agent needs to select one view for navigation at each step and identify a specific object among all candidate objects at the destination. However, most of the previous approaches fail to distinguish between views and objects, instead processing them using the vanilla vision encoder, which results in ambiguous representations of both views and objects. To address the above issues, we propose ViSMoE, which equips sparse Mixture-of-Experts with a visual-aware routing policy for the embodied agent. This framework processes different types of visual information specifically, resulting in discriminative visual representations for both views and objects. Experimental results on REVERIE and SOON datasets demonstrate that ViSMoE outperforms the previous state-of-the-art methods, showing the superiority of our proposed method.

cs.CV↗

arXiv2026

AirAlign: Geometry-Aware Relative Pose Alignment for UAV Last-Meter Navigation

Unmanned aerial vehicle (UAV) navigation in modern low-altitude environments requires more accurate pose alignment in the final approach stage for target information acquisition or manipulation, making "last-meter" navigation increasingly important. However, severe viewpoint and appearance variations make this task challenging. To tackle this problem, we propose AirAlign, a framework for RGB-only image-pair relative pose alignment for UAVs. AirAlign uses a pretrained visual geometry reconstruction model as the backbone to extract geometry-aware features from source-target image pairs. In addition, to better utilize the limited training data, we split the training set into multiple scene-disjoint folds for unseen cross-validation and model selection. During inference, the predictions of the selected models are averaged to form the ensemble output of the overall framework. Experiments on the PairUAV challenge at the ACMMM 2026 Workshop on UAVs in Multimedia demonstrate the effectiveness and robustness of our method, while comprehensive ablation studies validate the contribution of each component.

cs.CV↗

arXiv2026

DUD: Decoupled Update Dynamics for Reliable Uncertainty Quantification in Large Language Models

Accurate Uncertainty Quantification (UQ) is critical for reliable deployment of Large Language Models (LLMs), yet traditional probability-based metrics often fail to capture the model's true epistemic state. While recent mechanistic approaches leverage hidden state dynamics, they typically aggregate residual stream updates, conflating the distinct roles of parametric memory (Feed-Forward Networks) and contextual processing (Attention). We argue that this aggregation obscures fine-grained mechanistic conflicts, such as memory-context misalignment, that are fundamental indicators of uncertainty. To address this, we introduce \textbf{D}ecoupled \textbf{U}pdate \textbf{D}ynamics \textbf{(DUD)}, a framework that explicitly decouples FFN and Attention contributions via noise-induced causal interventions. By quantifying the independent restoration capabilities of each module, we construct a dual-stream dynamic profile that captures the model's internal fragility. Extensive experiments demonstrate that DUD significantly outperforms state-of-the-art baselines in both uncertainty estimation and calibration, while exhibiting superior cross-dataset generalization, validating decoupled dynamics as a robust proxy for model faithfulness.

cs.CL↗

arXiv2026

Diversity-Oriented Fine-Tuning for Uncertainty-Based Hallucination Detection

Existing hallucination detection methods are typically conducted at the inference stage, without making any modifications to the model itself. In this paper, we are interested in exploring fine-tuning strategies that enhance the detectability of hallucinations in the resulting model. Focusing on semantic-entropy-based detection, we observe that many erroneous outputs remain undetected because the model produces nearly identical incorrect answers across multiple runs. To address this, we propose diversity-oriented fine-tuning to encourage more varied generations. We introduce two specific strategies: one based on Supervised Fine-Tuning (SFT) and the other on Direct Preference Optimization (DPO). Extensive experiments are conducted to evaluate our approach and analyze the behavior of the models before and after fine-tuning. We find that after adopting our fine-tuning methods, the models become less likely to produce low semantic entropy responses for hallucinated answers, thereby improving the effectiveness of hallucination detection, eventually yielding results better than or comparable with state of the art methods. The code will be publicly released.

cs.AI↗

arXiv2026

DeLS-Spec: Decoupled Long-Short Contexts for Parallel Speculative Drafting

Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel. Block-parallel drafters such as DFlash further improve drafting efficiency by predicting an entire block in one pass, but their position-wise predictions lack explicit intra-block causal conditioning. Recent methods such as Domino and DSpark attempt to introduce such causality into block-parallel drafting, but they require training the draft model from scratch, which limits their flexibility and increases training cost. We propose DeLS-Spec, a decoupled long-short context speculative decoding method. DeLS-Spec treats the fixed DFlash model as a long-context expert and introduces a lightweight local head as a short-context expert. The local head can be trained independently with a standard next-token prediction objective, without joint training with the target model or the DFlash backbone, leading to extremely low training cost. At inference time, DeLS-Spec combines long-context and short-context logits, and the local head is not tied to a specific DFlash checkpoint, making the method more modular and flexible. Experiments on Qwen3 models show that DeLS-Spec consistently improves speedup and average acceptance length over DFlash across math, code, and dialogue benchmarks.

cs.CL↗

arXiv2026

ReFreeKV: Towards Threshold-Free KV Cache Compression

To reduce memory consumption during LLM inference, a handful of methods have been proposed for KV cache pruning. While these techniques can accomplish lossless memory reduction on many datasets, they often hinge on an under-emphasized condition: an input/domain-specific threshold for KV cache budget needs to be pre-determined to achieve the optimal performance. However, such input-sensitive design may be considerably limited in real-world scenarios, as open-domain inputs span diverse domains, lengths and difficulty levels, without clear boundaries for threshold selection. As a result, the dependence of such input-sensitive threshold can be a fundamental limitation that causes large degradation on arbitrary inputs. In this work, we propose a new objective that lifts the threshold constraints for robust KV compression, advocating for "threshold-free" methods that adaptively adjust budget allocation while preserving full-cache performance. We then propose a novel method, ReFreeKV, serving as the first instantiation of this objective. Extensive experiments across 13 datasets with diverse context lengths, task types, and model sizes demonstrate its efficacy and efficiency. Our code is publicly released at https://github.com/Patrick-Ni/ReFreeKV.

cs.CL↗

arXiv2026

Investigating Advanced Reasoning of Large Language Models via Black-Box Environment Interaction

Existing tasks fall short in evaluating reasoning ability of Large Language Models (LLMs) in an interactive, unknown environment. This deficiency leads to the isolated assessment of deductive, inductive, and abductive reasoning, neglecting the integrated reasoning process that is indispensable for human-like discovery learning. We introduce a novel evaluation paradigm, \textit{black-box environment interaction}, to tackle this challenge. A black-box environment is defined by a hidden function that maps a specific set of inputs to outputs. LLMs are required to unravel the hidden function behind the black-box environment by interacting with it in given exploration turns, and reasoning over observed input-output pairs. Leveraging this idea, we build the \textsc{Oracle} benchmark which comprises 6 types of black-box task with 96 black-box environments. 19 modern LLMs are benchmarked. o3, a leading LLM from OpenAI, ranks first in 5 of the 6 tasks, achieving over 70\% accuracy on most easy black-box environments. But it still struggles with some hard black-box tasks, where the average performance drops below 40\%. Further analysis reveals a universal difficulty among LLMs: They lack the high-level planning capability to develop efficient and adaptive exploration strategies for hypothesis refinement. Code is available in https://github.com/lemonsis/Oracle_Benchmark.

cs.AI↗

arXiv2026

CRISP: Compressing Redundancy in Chain-of-Thought via Intrinsic Saliency Pruning

Long Chain-of-Thought (CoT) reasoning is pivotal for the success of recent reasoning models but suffers from high computational overhead and latency. While prior works attempt to compress CoT via external compressor, they often fail to align with the model's internal reasoning dynamics, resulting in the loss of critical logical steps. This paper presents \textbf{C}ompressing \textbf{R}edundancy in Chain-of-Thought via \textbf{I}ntrinsic \textbf{S}aliency \textbf{P}runing (\textbf{CRISP}), a framework that compresses CoT by exploiting the model's intrinsic saliency. Our analysis reveals a distinct phenomenon: the reasoning termination token \texttt{[object Object]} acts as an information anchor, where its attention pattern effectively demarcates essential reasoning from redundancy. Based on this finding, we design a policy that utilizes these intrinsic attention signals to guide atomic compression operations. In contrast to coarse-grained pruning strategies, CRISP strategically distills the reasoning chain to maximize information density while preserving logical coherence. Empirical results across various backbone models and mathematical datasets demonstrate that CRISP achieves a 50-60% reduction in token count without compromising accuracy, effectively mitigating the efficiency bottleneck of long-context reasoning. We open-source our implementation to facilitate further research in efficient reasoning.

cs.CL↗

arXiv2026

Language Reconstruction with Brain Predictive Coding from fMRI Data

Many recent studies have shown that the perception of speech can be decoded from brain signals and subsequently reconstructed as continuous language. However, there is a lack of neurological basis for how the semantic information embedded within brain signals can be used more effectively to guide language reconstruction. Predictive coding theory suggests the human brain naturally engages in continuously predicting future words that span multiple timescales. This implies that the decoding of brain signals could potentially be associated with a predictable future. To explore the predictive coding theory within the context of language reconstruction, this paper proposes \textsc{PredFT}~(\textbf{F}MRI-to-\textbf{T}ext decoding with \textbf{Pred}ictive coding). \textsc{PredFT} consists of a main network and a side network. The side network obtains brain predictive representation from related regions of interest~(ROIs) with a self-attention module. The representation is then fused into the main network for continuous language decoding. Experiments on two naturalistic language comprehension fMRI datasets show that \textsc{PredFT} outperforms current decoding models on several evaluation metrics.

cs.CL↗

arXiv2026

AgentVLN: Towards Agentic Vision-and-Language Navigation

Vision-and-Language Navigation (VLN) requires an embodied agent to ground complex natural-language instructions into long-horizon navigation in unseen environments. While Vision-Language Models (VLMs) offer strong 2D semantic understanding, current VLN systems remain constrained by limited spatial perception, 2D-3D representation mismatch, and monocular scale ambiguity. In this paper, we propose AgentVLN, a novel and efficient embodied navigation framework that can be deployed on edge computing platforms. We formulate VLN as a Partially Observable Semi-Markov Decision Process (POSMDP) and introduce a VLM-as-Brain paradigm that decouples high-level semantic reasoning from perception and planning via a plug-and-play skill library. To resolve multi-level representation inconsistency, we design a cross-space representation mapping that projects perception-layer 3D topological waypoints into the image plane, yielding pixel-aligned visual prompts for the VLM. Building on this bridge, we integrate a context-aware self-correction and active exploration strategy to recover from occlusions and suppress error accumulation over long trajectories. To further address the spatial ambiguity of instructions in unstructured environments, we propose a Query-Driven Perceptual Chain-of-Thought (QD-PCoT) scheme, enabling the agent with the metacognitive ability to actively seek geometric depth information. Finally, we construct AgentVLN-Instruct, a large-scale instruction-tuning dataset with dynamic stage routing conditioned on target visibility. Extensive experiments show that AgentVLN consistently outperforms prior state-of-the-art methods (SOTA) on long-horizon VLN benchmarks, offering a practical paradigm for lightweight deployment of next-generation embodied navigation models. Code: https://github.com/Allenxinn/AgentVLN.

cs.RO↗

arXiv2026

Temporal Guidance for Large Language Models

Contrastive Decoding (CD) enhances the generation quality of large language models (LLMs) but incurs significant additional computational overhead due to the need for an auxiliary model. Existing internal self-contrastive decoding methods, such as Decoding by Contrasting Layers (DoLa), focus on discrepancies across different layers, which are notably unstable on small-scale models. In this work, based on the observation that LLMs exhibit local preferences, we propose a novel contrastive guidance strategy along the temporal dimension, namely Temporal Guidance (TeGu). Our method ingeniously leverages Multi-Token Prediction (MTP) to construct weaker amateur predictions for model self-contrast. To standardize the implementation of this mechanism, we further introduce a lightweight Conditional MTP Projector (cMTPP), which avoids maintaining multiple independent networks as required by other MTP modules. Across various model series and benchmarks, TeGu achieves significant performance improvements while maintaining low additional memory consumption and computational overhead.

cs.CL↗

arXiv2026

From Reasoning LLMs to BERT: A Two-Stage Distillation Framework for Search Relevance

Query-service relevance prediction in e-commerce search systems faces strict latency requirements that prevent the direct application of Large Language Models (LLMs). To bridge this gap, we propose a two-stage reasoning distillation framework to transfer reasoning capabilities from a powerful teacher LLM to a lightweight, deployment-friendly student model. In the first stage, we address the limitations of general-purpose LLMs by constructing a domain-adapted teacher model. This is achieved through a three-step process: domain-adaptive pre-training to inject platform knowledge, supervised fine-tuning to elicit reasoning skills, and preference optimization with a multi-dimensional reward model to ensure the generation of reliable and preference-aligned reasoning paths. This teacher can then automatically annotate massive query-service pairs from search logs with both relevance labels and reasoning chains. In the second stage, to address the challenges of architectural heterogeneity in standard distillation, we introduce Contrastive Reasoning Self-Distillation (CRSD). By modeling the behavior of the same student model under ``standard'' and ``reasoning-augmented'' inputs as a teacher-student relationship, CRSD enables the lightweight model to internalize the teacher's complex decision-making mechanisms without needing the explicit reasoning path at inference. Offline evaluations and online A/B testing in the Meituan search advertising system demonstrate that our framework achieves significant improvements across multiple metrics, validating its effectiveness and practical value.

cs.IR↗

arXiv2026

Hallucination Mitigating for Medical Report Generation

In the realm of medical report generation (MRG), the integration of natural language processing has emerged as a vital tool to alleviate the workload of radiologists. Despite the impressive capabilities demonstrated by large vision language models (LVLMs) in understanding natural language, their susceptibility to generating plausible yet inaccurate claims, known as ``hallucinations'', raises concerns-especially in the nuanced and critical field of medical. In this work, we introduce a framework, \textbf{K}nowledge-\textbf{E}nhanced with Fine-Grained \textbf{R}einforced Rewards \textbf{M}edical Report Generation (KERM), to tackle the issue. Our approach refines the input to the LVLM by first utilizing MedCLIP for knowledge retrieval, incorporating relevant lesion fact sentences from a curated knowledge corpus. We then introduce a novel purification module to ensure the retrieved knowledge is contextually relevant to the patient's clinical context. Subsequently, we employ fine-grained rewards to guide these models in generating highly supportive and clinically relevant descriptions, ensuring the alignment of model's outputs with desired behaviors. Experimental results on IU-Xray and MIMIC-CXR datasets validate the effectiveness of our approach in mitigating hallucinations and enhancing report quality.

cs.CL↗

arXiv2025

Retrieval-augmented Prompt Learning for Pre-trained Foundation Models

The pre-trained foundation models (PFMs) have become essential for facilitating large-scale multimodal learning. Researchers have effectively employed the ``pre-train, prompt, and predict'' paradigm through prompt learning to induce improved few-shot performance. However, prompt learning approaches for PFMs still follow a parametric learning paradigm. As such, the stability of generalization in memorization and rote learning can be compromised. More specifically, conventional prompt learning might face difficulties in fully utilizing atypical instances and avoiding overfitting to shallow patterns with limited data during the process of fully-supervised training. To overcome these constraints, we present our approach, named RetroPrompt, which aims to achieve a balance between memorization and generalization by decoupling knowledge from mere memorization. Unlike traditional prompting methods, RetroPrompt leverages a publicly accessible knowledge base generated from the training data and incorporates a retrieval mechanism throughout the input, training, and inference stages. This enables the model to actively retrieve relevant contextual information from the corpus, thereby enhancing the available cues. We conduct comprehensive experiments on a variety of datasets across natural language processing and computer vision tasks to demonstrate the superior performance of our proposed approach, RetroPrompt, in both zero-shot and few-shot scenarios. Through detailed analysis of memorization patterns, we observe that RetroPrompt effectively reduces the reliance on rote memorization, leading to enhanced generalization.

cs.CL↗

arXiv2025

LSRS: Latent Scale Rejection Sampling for Visual Autoregressive Modeling

Visual Autoregressive (VAR) modeling approach for image generation proposes autoregressive processing across hierarchical scales, decoding multiple tokens per scale in parallel. This method achieves high-quality generation while accelerating synthesis. However, parallel token sampling within a scale may lead to structural errors, resulting in suboptimal generated images. To mitigate this, we propose Latent Scale Rejection Sampling (LSRS), a method that progressively refines token maps in the latent scale during inference to enhance VAR models. Our method uses a lightweight scoring model to evaluate multiple candidate token maps sampled at each scale, selecting the high-quality map to guide subsequent scale generation. By prioritizing early scales critical for structural coherence, LSRS effectively mitigates autoregressive error accumulation while maintaining computational efficiency. Experiments demonstrate that LSRS significantly improves VAR's generation quality with minimal additional computational overhead. For the VAR-d30 model, LSRS increases the inference time by merely 1% while reducing its FID score from 1.95 to 1.78. When the inference time is increased by 15%, the FID score can be further reduced to 1.66. LSRS offers an efficient test-time scaling solution for enhancing VAR-based generation.

cs.CV↗