跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+0新增论文
+0其中 AI 相关
+0已有论文修订
+92新增 AI 导读

arXiv 最近完成扫描:2026-11-23 01:07 · 全站最近入库:2026-11-21 12:19各信源进度 →

统计口径与数据新鲜度

统计 2026-11-22 07:05 — 2026-11-23 07:05(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-23 01:07最近发现新增版本:2026-11-21 12:19库内论文最新更新时间:2026-11-20 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

29,158 已收录论文24,181 AI 相关8,992 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 6,576 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-23 01:05(北京时间)
版本更新 · v2arXiv:2607.11012

EasyOPD: An Easy-to-use On-Policy Distillation Framework for Large Language Models

Jie Sun, Mao Zheng, Mingyang Song, Qiyong Zhong, Gengsheng Li, Zhepei Hong, Chang Wu, Pengfei Liu 等 10 位作者

原文摘要

Conventional language-model distillation often relies on fixed teacher-generated data, which may not cover the states encountered by an evolving student policy. On-policy distillation (OPD) instead collects teacher or evaluator supervision on student-generated rollouts. However, existing OPD methods differ substantially in supervision form, tokenizer compatibility, teacher access, and supervision granularity, leading to fragmented implementations that are difficult to reproduce and extend. We present \textsc{EasyOPD}, an on-policy distillation framework built on verl, a distributed reinforcement-learning framework for large language models. \textsc{EasyOPD} separates user-side configuration, method-specific supervision logic, and verl-based execution. Its method modules connect to the shared backend through extension boundaries for loss construction, rollout metadata, reward processing, tokenizer alignment, and teacher-side computation. We instantiate representative methods for three OPD settings -- cross-tokenizer OPD, on-policy self-distillation, and step-wise OPD. Experiments on reasoning, code-generation, scientific-knowledge, and tool-use benchmarks show that these implementations can be executed through the same verl-based backend while retaining their method-specific objectives and task-dependent performance profiles. We release \textsc{EasyOPD} with runnable YAML configurations, documentation, and an installable demonstration package and video.

cs.CL

新论文 · v1arXiv:2611.24086

Constructing Dynamic Conversations with Neural Back-channel Generation and Voice-line Scheduling

Yifang Pan, Chris Landreth, Karan Singh

原文摘要

We present a novel approach to the problem of combining two voice-line channels of dyadic conversational audio into a unified conversation with controllable back-channeling and turn-taking dynamics. In voice performances for the game industry, the two tracks of voice-lines are typically recorded or synthesized in isolation. Our method takes two such unsynchronized tracks and uses a neural scheduler to sequence them into a single dynamic conversation, modifying both tracks by inserting or removing pauses and introducing voice overlap as appropriate. We further propose a back-channel predictor that is able to predict opportunity for back-channels for a given conversational turn's audio and content, which the scheduler can then incorporate. We evaluate our scheduler and predictor independently and in combination by: various quantitative comparison; a qualitative and perceptual comparison to prior art and naive baselines; a compelling set of results showing animator control over the diversity, extent and nature of back-channeling and turn dynamics.

cs.SD · cs.CL

版本更新 · v4arXiv:2609.23697

Distill What You Trust: Reliability-Aware Multi-Teacher On-Policy Distillation

Jie Sun, Mao Zheng, Mingyang Song, Zeyuan Liu, Gengsheng Li, Houcheng Jiang, Yilin Cheng, Bichuan Feng 等 11 位作者

原文摘要

Multi-teacher on-policy distillation allows a student to learn from complementary specialists on its own trajectories. Domain-routed approaches, however, select one teacher per example and keep it fixed throughout the response. This design both depends on domain labels that mixed training corpora often lack and cannot adapt teacher selection when the expertise required changes within a trajectory. We observe that each specialist deviates more from a shared reference on in-domain prompts than on out-of-domain prompts, on average. Based on this observation, we propose \textbf{TrustMOPD}, which replaces example-level teacher selection with label-free, token-level supervision allocation. At each student-generated prefix, TrustMOPD measures this displacement in next-token preferences, calibrates its magnitude across teachers, and uses the resulting scores as proxies for local reliability to weight teacher-specific distillation losses. Evaluated across mathematics, code, and instruction following, TrustMOPD closes 91.5\% and 98.0\% of the overall-score gap between the initial student and oracle-routed teachers when trained on \textsc{SingleCap} and \textsc{MultiCap}, respectively, compared with 54.4\% and 54.5\% for the strongest label-free baseline in each setting. On \textsc{SingleCap}, it approaches label-based MOPD without using domain labels. Our code is available at https://github.com/sunjie279/TrustMOPD-.

cs.CL · cs.LG

新论文 · v1arXiv:2611.24076

Meaning Reaches Object Choice but Not Completion: Where Is a VLA Vulnerable to Landmark Synonyms?

Ryoichi Kurihara, Kota Sunamoto

原文摘要

VLA policies are fragile to rewording. Success rates alone cannot say whether a policy fails to choose the object or to complete the task, nor where in the policy to intervene. We study the released GR00T N1.7 on two LIBERO-Spatial tasks differing only in the landmark word telling which of two bowls to take. Synonyms absent from the training instructions send the policy to the bowl they name, unlike made-up words, yet complete 67.6% of one task's closed-loop episodes, against the original word's 100%. Before the connector between the vision-language model and the action generator, replacing only the landmark tokens' vectors with the other landmark's (pasting) sends nearly every episode to the other bowl. After the connector, pasting at the landmark, the remaining instruction, or the image tokens alone rarely moves the choice, but any two do. Hiding the landmark tokens from the action generator after the connector still lowers completion for the synonyms that fall short. Finetuning on diversified landmark words repairs that one task's completion for synonyms absent from finetuning; finetuning on the original instructions does not. Training only the connector on the diversified words suffices, as does copying the connector of a policy fully finetuned on them into one finetuned on the original instructions. An open-loop score on starting frames misses that weights outside the connector also carry part of the repair. For now, on this evidence, robustness to rewording should be judged in closed loop, with actions executed, intervening on both sides of the connector, scoring choice and completion separately.

cs.RO · cs.CL · cs.LG

新论文 · v1arXiv:2611.24071

Measuring and Charging for Retained Context in Long-Context LLM Services

Yi Lu, Qianshi Lin, Yumeng Yang, Hongxia Wang

原文摘要

Long-context large language model (LLM) services charge for repeated context, yet billing weights need not reflect its contribution to answer quality. We present a three-layer framework separating content relevance, observable proxies and auditable positional weights. Context-ablation experiments show that relevance dominates position and that distance-decay families do not outperform a relevance-only baseline. A measured serving stack shows that cache hits reduce repeated-prefix compute cost by over two orders of magnitude at 65,536 tokens, while cold resends constrain sustainable discounts. Cross-provider measurements separate cache value from tokenisation and list-price differences. Calibrated simulations reduce the 99th percentile of session charges by 31-42%, although revenue-matched nonlinear tariffs achieve similar reductions and median charges increase. The framework provides a bounded per-token accounting rule and measured deployment constraints, rather than a validated utility estimator.

cs.PF · cs.CL

新论文 · v1arXiv:2611.24058

MIRAGE: A Framework for Multi-hop Reasoning under Dynamic Knowledge Injection

Heyang Xu, HanWen Hao, Haifeng Wang, Haoyi Zhou, Jianxin Li

原文摘要

Multi-hop Question Answering (MQA) requires large language models (LLMs) to retrieve and synthesize multiple pieces of evidence into a coherent reasoning path, which is essential for complex tasks. However, prior work on knowledge updating for MQA has largely focused on simple fact edits, overlooking a more challenging and realistic setting---\emph{Knowledge Injection}---in which the model must simultaneously accommodate both conflicting updates and entirely novel information. Under this setting, a hop-level answer is often wrong simply because the model trusts the wrong knowledge source: it clings to outdated parametric memory when the injected knowledge conflicts with it (K2), and it cannot answer when the required fact is entirely new (K3). To investigate this, we introduce NEXUS, the first benchmark tailored for multi-hop reasoning under Knowledge Injection, which explicitly covers both conflicting (K2) and incremental (K3) knowledge. We further propose MIRAGE, an efficient, training-free framework whose core component, the Correlation-aware Dual-path Reasoner (CDR), resolves such per-hop source-trust errors via P(True)-based confidence routing, complemented by a lightweight template-guided corrector for residual structural deviations. Extensive experiments across four LLM families and two benchmarks demonstrate that MIRAGE consistently improves both final-answer accuracy and hop-level reasoning accuracy, establishing a new state of the art in this challenging setting.

cs.CL

新论文 · v1arXiv:2611.24055

Grading Executable Legal Rules Against Independent Records: Transitional Provisions and Korean SME Status

Jinhyung Bae

原文摘要

Laws are increasingly turned into programs that decide cases, such as whether a firm was a small or medium enterprise (SME) on a given date, and large language models (LLMs) now write them. The programs are usually graded against answers written by someone who read the same law, so a program can score well by sharing the grader's reading rather than by getting the law right. LexTemporis grades them instead against records of what actually happened: 10,901 firm-years of SME status disclosed by Korean listed companies, with every version of the governing rules. An appendix adds comparison cases from US income tax and German hospital minimum volumes. The date changes outcomes mainly through provisions outside the main statutory text: one transitional provision raises agreement with the disclosures from 83-86% to 89-90%. The disclosures also test the author's reading of a contested provision, favouring it in 130 of 189 firm-years where readings diverge. Four LLMs wrote the rules, given different selections of the law. A pre-registered comparison finds no condition better across all four models. In exploratory analyses, supplying the right provisions looks better graded against the author's reading than against the records: the advantage shrinks for three models (for one, to between a quarter and three-fifths of its size), and per firm-year it is several times larger where the author's reading is wrong, partly by construction, since there reproducing that reading more faithfully must lose. Grading legal programs against the grader's own reading can overstate how well they work.

cs.CL · cs.AI

版本更新 · v2arXiv:2611.21519

Embedding Cosine Measures Text Length Before It Measures Homogenization

Fengming Liu, Mingyu Shu

原文摘要

Mean pairwise cosine similarity between transformer embeddings is a standard measure of whether online discourse is converging under generative AI. When the embeddings are mean-pooled outputs of an encoder never trained for sentence similarity, this measure mostly tracks text length. In 6.06 million Reddit comments, an uncentered DeBERTa-v3 measure had suggested that only conservative discourse broke toward homogeneity after late 2022. Truncating the same comments from 10 to 150 words raises their mean cosine by eight times the reported break. A placebo that swaps every comment for a random one of equal length reproduces 81% of the asymmetry, and a fixed-content simulation reproduces 78%. The leading principal direction carries 99.9% of raw DeBERTa homogeneity, and once length is handled the asymmetry disappears under every encoder. On 575,000 Stack Exchange answers the measure reports a post-ChatGPT homogenization of technical sites that the placebo fully reproduces, although answers grew longer. A census of 31 studies locates the risk where anisotropic encoders meet time-series or length-unequal designs. A reverse control on a well-known experiment shows that centering with renormalization can reverse a genuine effect. We propose a low-cost diagnostic protocol. Corrected measures show that the two political camps converged between 2019 and 2022 and stopped converging before ChatGPT. Cross-camp replies engage their targets more closely than same-camp replies.

cs.CL · cs.CY · cs.SI

新论文 · v1arXiv:2611.24049

A Hybrid RAG Framework for Deceptive Review Detection: Integrating RoBERTa-LoRA, FAISS-Based Retrieval, and LLM Reasoning

Andrei Kostin, José Ángel Diaz-Garcia, Karel Gutiérrez-Batista

原文摘要

The growth of deceptive online opinions poses a critical risk of misinformation and fraud. This paper proposes a Retrieval-Augmented Generation (RAG) system to automatically distinguish spam reviews. The architecture integrates RoBERTa with Low-Rank Adaptation (LoRA) classification, FAISS semantic vector retrieval, and a Large Language Model (LLM) for final reasoning. The system was evaluated across balanced and severely imbalanced real-world scenarios. Results demonstrate that five epochs of fine-tuning enable robust domain transfer, achieving an F1-score of 0.90 in cross-domain evaluation, outperforming baselines (OpenAI Detector,F1 = 0.82) and demonstrating competitive performance against specialized in-domain SOTA models (fakeRoBERTa,F1 = 0.97). The proposed architecture provides evidence-backed explainability and statistically significant precision gains while maintaining comparable overall detection performance.

cs.CL

版本更新 · v2arXiv:2611.23158

MoE--LoRA Methods for Multilingual Adaptation: A Systematic Empirical Study

Nikit Srivastava, Joel Dag, René Speck, Axel-Cyrille Ngonga Ngomo

原文摘要

Routed low-rank adapters offer a parameter-efficient means to improve the performance of large language models (LLMs) across tasks. Hence, it is assumed that they can also be used to improve the support of low-resource languages in LLMs, which currently serve English far better than the thousands of other languages spoken worldwide. However, this assumption has never been verified empirically under a shared multilingual evaluation protocol. We address exactly this gap by comparing eight low-rank adaptation (LoRA) methods on a frozen base model, using 146 training runs across 200 unique languages and more than 10,000 evaluations. Our results suggest that a reduced-scale search can identify viable configurations of LoRA methods but does not reliably predict their relative performance, as the top-ranking configuration is never best at scale. LoRA reliably improves language-model fit on translation data, consistently degrades reading comprehension, and benefits low-resource languages most. These findings suggest that LoRA methods should be regarded as targeted tools rather than general-purpose means for improving the multilingual capabilities of LLMs.

cs.CL · cs.LG

新论文 · v1arXiv:2611.24043

Compression Effects on Low-Resource Multilingual Translation

Abbas Sani

原文摘要

We present a pilot study characterising the effect of post-training quantisation on multilingual machine translation quality and inference cost for three African languages - Hausa, Swahili, and Yoruba - across three model families: a small general-purpose LLM (Qwen2.5-1.5B), a larger multilingual instruction-tuned LLM (Aya-23-8B), and a dedicated multilingual MT model (NLLB-200-distilled-1.3B). Evaluating on the FLORES-200 devtest set (1,012 sentences per direction) under FP16, bitsandbytes INT8, and NF4 INT4 precisions on consumer-grade NVIDIA T4 GPUs, we find: (i) the specialised MT model outperforms both LLMs by 32.7-43.3 chrF on the Hausa and Swahili directions, and by 12.9-24.7 chrF on Yoruba, despite being roughly a sixth the size of the 8B instruction-tuned LLM; (ii) post-training quantisation preserves aggregate translation quality within approximately 1 chrF across all model families and language directions, with a single worst case of 2.19 chrF; (iii) on consumer GPUs, INT8 and INT4 quantisation provide substantial memory savings (49-61% at INT4) but do not reduce inference latency relative to FP16 - in fact, latency increases by 1.3x to 5.3x; (iv) a clear performance gradient tracking training-resource availability is visible across languages (Swahili > Hausa > Yoruba). These findings motivate a research agenda focused on hardware-aware compression and error-type analysis beyond aggregate metrics for low-resource deployment.

cs.CL

版本更新 · v2arXiv:2505.07460

A Survey on Collaborative Mechanisms Between Large and Small Language Models

Yi Chen, Haiquan Wang, JiaHao Zhao, HaoHao Han

原文摘要

Large Language Models (LLMs) deliver powerful AI capabilities but face deployment challenges due to high resource costs and latency, whereas Small Language Models (SLMs) offer efficiency and deployability at the cost of reduced performance. Collaboration between LLMs and SLMs emerges as a crucial paradigm to synergistically balance these trade-offs, enabling advanced AI applications, especially on resource-constrained edge devices. This survey provides a comprehensive overview of LLM-SLM collaboration, detailing various interaction mechanisms (pipeline, routing, auxiliary, distillation, fusion), key enabling technologies, and diverse application scenarios driven by on-device needs like low latency, privacy, personalization, and offline operation. While highlighting the significant potential for creating more efficient, adaptable, and accessible AI, we also discuss persistent challenges including system overhead, inter-model consistency, robust task allocation, evaluation complexity, and security/privacy concerns. Future directions point towards more intelligent adaptive frameworks, deeper model fusion, and expansion into multimodal and embodied AI, positioning LLM-SLM collaboration as a key driver for the next generation of practical and ubiquitous artificial intelligence.

cs.AI · cs.CL

版本更新 · v2arXiv:2609.24657

Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models

Xiaoqiang Wang, Mengyang Xiong, Jun Dai, Bang Liu

原文摘要

Improving pretrained language models with limited adaptation data calls for computational modules that are both effective and scalable. Parameterized quantum circuits offer expressive transformations for language modeling, yet existing quantum-enhanced language models apply the same predefined circuit design to every token in an input sequence, regardless of its content. This limitation is particularly acute in masked-diffusion language models, where token representations evolve across iterative denoising steps while the circuit design remains fixed. We introduce HyperQ, a quantum-augmented masked-diffusion language model that dynamically synthesizes token-specific quantum circuits. Built upon an IQP-inspired architecture, HyperQ employs a lightweight circuit hypernetwork to generate rotation angles and measurement bases conditioned on each token's hidden state. The resulting quantum expectation values are projected back into the frozen language-model backbone through a residual connection. To support wider quantum registers, HyperQ bounds circuit interaction degree, maintaining per-qubit gradient scale and readout cost independent of register width. Across six zero-shot benchmarks, scaling HyperQ from 16 to 64 qubits improves average accuracy from 47.65 to 54.30, surpassing the frozen backbone's 49.59 by 4.71 points while using only one tenth of the fine-tuning data used by classical baselines. We further execute the generated circuits on an IBM 156-qubit superconducting quantum processor, demonstrating their feasibility on physical quantum hardware.

quant-ph · cs.CL

新论文 · v1arXiv:2611.24001

Rethinking Fixed-Horizon Latent Reasoning

Tianxiang Chen, Jingyuan Zhou, Weiming Yang, Kaidi Yang

原文摘要

Existing latent reasoners typically run for a preset latent horizon during inference. Intuitively, a longer horizon should provide more opportunity for reasoning. However, in this work, we find that extending latent Chain-of-Thought (CoT) to the full preset horizon may in fact degrade reasoning performance. We observe that latent CoT evolves through distinct functional stages, with earlier computation forming frontier reasoning states, while later computation mainly reorganizes existing candidates and can weaken gold candidate support. These findings motivate a rethink of fixed-horizon latent inference. To address this, we characterize the reasoning benefits of latent CoT by realized and remaining constructive progress. Empirically, we find that higher realized constructive progress corresponds to higher current reasoning accuracy, while more remaining progress corresponds to larger future gains from continued latent CoT. To this end, we introduce a lightweight method that estimates the remaining reasoning progress and adaptively adjusts the latent horizon during inference. Across representative latent CoT reasoners and multiple arithmetic benchmarks, our method improves reasoning accuracy by up to 7.2 percentage points while reducing latent computation by up to 29.0%, without retraining the LLM.

cs.LG · cs.AI · cs.CL

新论文 · v1arXiv:2611.24000

A Mechanistic Analysis of In-Context Learning and Out-of-Context Hallucination

Tianxiang Chen, Jingyuan Zhou, Guoyao Yang, Kaidi Yang

原文摘要

Large language models can infer and follow specific rules through in-context learning (ICL), yet may fail to remain faithful to provided evidence, giving rise to out-of-context (OOC) hallucination. Existing work has largely investigated the underlying mechanisms of ICL and OOC hallucination along separate lines. However, this separation obscures the mechanistic basis for these seemingly opposing outcomes. In this work, we study this contrast through the unified view of functional roles in dynamic context. The key observation is that role information can be internally represented without reliably constraining output. We derive a first-order decomposition to characterize the role effect, showing that its substantial influence depends on two necessary conditions, namely sufficient contribution magnitude and coherent signed aggregation. Causal analysis further shows that role span contributions affect key token choices involving response structure, content selection, and factual values. We further show that role-consistent choices depend on sufficient role span support at these key decisions, while failing choices can be reversed when this support overcomes the deficit. These findings support a common mechanistic view of ICL and OOC hallucination: each functional role defines the pattern of output decisions over which role span contributions must remain effective, and differences in this organization account for the seemingly opposing outcomes.

cs.CL · cs.AI · cs.LG

新论文 · v1arXiv:2611.23969

ReflectRec: Cycle-Consistency Preference Verification for LLM-based Recommendation

Dae Kyoo Kim, Heejung Cho, Sungmoo Hong, Eunil Park

原文摘要

CoT-style LLM recommenders generate a natural-language preference note as a chain-of-thought prefix to ranking, but the note is trusted without verification, and the ranker never consults behaviorally similar users. We propose ReflectRec, organized around a preference-to-history round-trip verifier whose recall score serves as an operational signal of preference-to-history recoverability, enabling conditional regeneration for low-recoverability preferences and selection among regeneration attempts. Around this verifier, we wire a semantic Stage-1 prompt, conditional regeneration, and Forward-Look retrieval that injects future watches of behaviorally similar users as a collaborative hint. On MovieLens-100K and MovieLens-1M, the full pipeline outperforms CoT4Rec-ICL by +16.25 / +4.46 pp Hit@10, with positive gains across all ValScore buckets and three LLM backbones (Qwen3-30B, Qwen2.5-72B, and Llama-3.3-70B). We also observe a similar verifier-guided regeneration pattern on Amazon Electronics.

cs.CL · cs.IR

版本更新 · v2arXiv:2602.10138

Multimodal Information Fusion for Chart Understanding: A Survey of MLLMs -- Evolution, Limitations, and Cognitive Enhancement

Zhihang Yi, Jian Zhao, Jiancheng Lv, Tao Wang

原文摘要

Chart understanding is a quintessential information fusion task, requiring the seamless integration of graphical and textual data to extract meaning. The advent of Multimodal Large Language Models (MLLMs) has revolutionized this domain, yet the landscape of MLLM-based chart analysis remains fragmented and lacks systematic organization. This survey provides a comprehensive roadmap of this nascent frontier by structuring the domain's core components. We begin by analyzing the fundamental challenges of fusing visual and linguistic information in charts. We then categorize downstream tasks and datasets, introducing a novel taxonomy of canonical and non-canonical benchmarks to highlight the field's expanding scope. Subsequently, we present a comprehensive evolution of methodologies, tracing the progression from classic deep learning techniques to state-of-the-art MLLM paradigms that leverage sophisticated fusion strategies. By critically examining the limitations of current models, particularly their perceptual and reasoning deficits, we identify promising future directions, including advanced alignment techniques and reinforcement learning for cognitive enhancement. This survey aims to equip researchers and practitioners with a structured understanding of how MLLMs are transforming chart information fusion and to catalyze progress toward more robust and reliable systems.

cs.CV · cs.AI · cs.CL · cs.LG

版本更新 · v3arXiv:2602.16169

Unlocking Continuous Diffusion Language Models via Discrete Stochastic Localization

Yunshu Wu, Jiayi Cheng, Longxuan Yu, Partha Thakuria, Rob Brekelmans, Evangelos E. Papalexakis, Greg Ver Steeg

原文摘要

Continuous diffusion is a natural framework for non-autoregressive generation but has generally lagged behind masked discrete diffusion models (MDMs) on discrete sequence generation. We argue that continuous diffusion can effectively model discrete sequences through the joint design of noise parameterization and token representation geometry. We introduce \emph{Discrete Stochastic Localization} (DSL), a continuous-state framework that combines a stochastic localization channel with unit-sphere token embeddings, resulting in a Bayes-optimal denoiser that depends only on the noisy state, without additional timestep conditioning. A single trained denoising network then supports multiple sampling methods, from continuous denoising to masked diffusion, random-order autoregressive generation and hybrid continuous-then-discrete generation. Parallel generation in this continuous-state framework, therefore, need not imply synchronized denoising: some tokens can establish context earlier while others remain open to refinement. Fine-tuning a pretrained MDLM checkpoint with DSL substantially improves generation quality, as measured by MAUVE, on OpenWebText across all step budgets from $T{=}128$ to $T{=}1024$. The same checkpoint supports random-order autoregressive sampling, and a hybrid continuous-then-discrete sampler using as few as $T=48$ total steps, without distillation or further training.

cs.LG · cs.CL

版本更新 · v2arXiv:2606.29959

Know Before You Fetch: When Does Confidence-Based Retrieval-Budget Allocation Pay Off?

Zhe Dong, Fang Qin, Manish Shah

原文摘要

Adaptive retrieval-augmented generation uses a signal from the reader, such as its confidence when answering without retrieval (closed-book), to decide whether and how much to retrieve. We compare such methods not with always retrieving many passages but with the best simple policy under the same cap on average passages per query: a fixed depth or a random mixture of fixed depths. Policies allocate zero, one or five passages and are scored on held-out questions. Seven settings of about 2,000 questions combine TriviaQA, Natural Questions and MS MARCO, Qwen3-8B and Qwen3-32B readers, and two retrieval sources: per-question passage pools and a 3.5-million-passage Wikipedia index. Allocation by confidence could gain in three ways. It rations a scarce budget: at a cap of half a passage per query it beats the best simple policy in all seven settings, by 0.010 to 0.075 in accuracy. It avoids harmful retrieval: at caps of one passage or more, substantial gains appear only on Wikipedia, where passages often overwrite known answers, most robustly for Qwen3-8B on TriviaQA (+0.051, also under exact match and F1). It barely helps to choose depth: closed-book confidence ranks which queries a first passage helps but ranks the choice between one and five passages little better than chance, and with passage pools a fixed depth of two or three passages is as good or better. Once the probe's latency is charged, gains remain only where the gate skips retrieval for most queries; for the larger reader they also depend on the answer metric.

cs.IR · cs.CL

版本更新 · v3arXiv:2610.16352

LoongSage: A Production-Grade, High-Efficiency Agentic RL Framework for Frontier LLMs with Validated Recipes

LoongSage Team, Bowei Song, Caiwei Zhao, Chuwen Liu, Dechuang Gu, Dong Fang, Feiyu Xie, Huaiyu Li 等 48 位作者

原文摘要

Reinforcement learning (RL) has emerged as a critical phase in the post-training pipeline of large language models (LLMs). As the frontier of LLM capabilities shifts from static reasoning to agentic tasks--characterized by external tool invocation, multi-turn interactions, and long-horizon decision-making--existing mainstream RL frameworks have grown increasingly unwieldy, raising the overhead of adoption, usage, and debugging. To address these challenges, we present LoongSage, an agent-native, fully asynchronous RL framework with a lightweight core of roughly 14K lines of code. Built upon Megatron for distributed training and SGLang for high-throughput inference, LoongSage centralizes orchestration within a single controller process and encapsulates all GPU-resident components as Ray actors. It introduces AgentFlow, a unified execution paradigm that orchestrates both white-box and black-box agents. For asynchronous execution, it integrates sliding-window strategies that bound trajectory staleness, and employs TransferMesh, a topology-aware synchronization mechanism that routes weights after each step over the fastest transport available to each receiver. Beyond agentic RL, LoongSage incorporates a memory-efficient Teacher Manager that decouples the scaling of teacher models from GPU memory constraints, natively supporting Multi-Teacher On-Policy Distillation. Extensive evaluations across four complete training runs, spanning complex reinforcement learning and multi-teacher distillation, demonstrate that task rewards improve steadily while the train-inference gap is kept under continuous control, and the fully asynchronous pipeline removes synchronous idle bubbles, sustaining hardware utilization.

cs.LG · cs.CL

版本更新 · v5arXiv:2607.23624

Where Is the Cost of Third-Party API Routers in Agentic Software Development?

Donghao Fu, Jingxin Li, Xue Jiang, Yihong Dong

原文摘要

Third-party API routers have become a common layer that unifies access across increasingly diverse LLM providers. In coding-agent workflows, high-autonomy operation is widely adopted because it reduces interaction overhead. As a result, a third-party API router, which sits between the agent and the upstream provider, inevitably occupies the trusted path. It can inspect and modify every request and response, yet no mechanism verifies alignment between the provider's output and the repository-level actions ultimately executed by the agent. Consequently, client-side permission mechanisms may become ineffective in practice. Whether this control gap produces real, hard-to-detect effects on software development tasks remains empirically unmeasured. In this paper, we conduct an empirical study of router-side injection in coding agents, examining four intervention levels of increasing subtlety: Response Substitution (L1), Response Append (L2), LLM-Polished Injection (L3), and LLM-Polished with Distribution Alignment Injection (L4). Moreover, we develop SIDEL, a framework for trace recording, replay, injection, and defense evaluation, with a curated dataset of 400 samples. We evaluate four representative coding agents, and further evaluate whitelist-based execution control and LLM review. Router-side intervention substantially alters repository-level actions and remains difficult for existing client-side safeguards to detect. Without additional mitigations, all evaluated agents achieved a defense success rate of 0 percent across all injection levels. Client-side mitigations and reactive reviews improve resistance but do not fully restore end-to-end control, motivating provider-side output-integrity guarantees. Our code is available at https://github.com/Riyasushin/SIDEL.

cs.SE · cs.AI · cs.CL

版本更新 · v3arXiv:2606.30852

When Does Learning to Stop Help? Configuring Risk-Controlled Early Exit for Reasoning-Model Services

Zhe Dong, Fang Qin, Manish Shah

原文摘要

Early exit lowers the cost of serving a reasoning model by stopping its reasoning once a provisional answer looks final. An operator must choose between a learned stopping rule and one static token budget, and savings against full reasoning do not settle the choice. We present a procedure that decides on calibration data from the workload. It certifies both families with Learn-then-Test to one bound on the fraction of all questions that full reasoning answers correctly but early exit does not, prices provisional answers as the serving stack does, and deploys the cheapest certified candidate. Across four models and five benchmarks neither family is better throughout. At a 5% risk target the learned rule saves 10.4-17.0 percentage points more tokens than the static budget on free-form mathematics with Qwen3, with accuracy 2.4-2.9 points below full reasoning, and less on MMLU-Pro on average. Which family costs less also changes with the risk target, in 8 of 15 settings, and with the price of a provisional answer. Across three prices, committing to one family in advance gives up as much as 27.4 points on some workload, whereas selecting on calibration data falls at most 11.6 points short of the better family, on a task with only 99 calibration questions. In a vLLM engine with Qwen3-8B, mean response time fell in proportion to generated tokens. Policies served on a task other than their calibration task exceeded the risk target in 18 of 44 task pairs; recalibration on the new task restored risk control.

cs.AI · cs.CL · cs.LG

新论文 · v1arXiv:2611.23900

Coverage Governs Accuracy: When to Expand the Lexicon and When to Fall Back in Low-Resource Manchu OCR

Ruixin Yu, Zifan Wang, Mingxin Ding, Xiaoxin Sun, Dongjie Wang, Suhua Wang, Minghao Yin

原文摘要

Character-level OCR for low-resource Manchu often produces near-miss spelling errors. Lexicon-constrained decoding can correct such errors with lexical knowledge, but a fixed dictionary cannot cover the many unseen word forms produced by agglutinative word formation. We study whether morphological expansion alleviates this limitation, and how the expansion size and confidence fallback affect recognition. We segment the training transcriptions with a dictionary-based stem list, extract roots and suffix chains from them, generate about 510k candidates, select the expansion size through frequency ranking on the validation set, and decompose accuracy under hard lexicon constraints into the product of coverage and in-lexicon precision. Across three CRNN training seeds, adding 250k candidates to a base dictionary of about 29k words raises the type accuracy on unseen combinations of seen roots from 63.6% to 86.3%, and overall token accuracy from 82.4% to 84.2%. Adding all candidates lowers overall token accuracy to 81.5%, showing that the coverage gain comes with a loss of in-lexicon precision. A further comparison shows that the base dictionary with confidence fallback performs best overall among the CRNN systems tested, reaching 76.1% type and 85.5% token accuracy and outperforming the expanded lexicon both with and without fallback. A Transformer recognizer shows similar trends. These results indicate that morphological expansion relieves the constraint of a fixed dictionary, but a larger lexicon does not guarantee better recognition; in the present low-resource Manchu OCR setting, a selective combination of reliable lexical constraints and open recognition is more effective.

cs.CL · cs.CV

版本更新 · v2arXiv:2611.21478

Jetriever: Shifting Agent Memory Retrieval from Similarity to Utility

Ruizhe Li, Licheng Zhang, Mingxuan Du, Zhendong Mao

原文摘要

Whether a stored memory helps answer a user's question often depends on world knowledge rather than surface similarity. Jetriever replaces similarity search with full-pool utility scoring: given a query, it uses Jev, an LLM-derived System One decision model, to score every stored memory for the probability that it is useful to the answer, ranking the memory pool directly by those probabilities. On InMind's 125 indirect-query tasks, Jetriever places the gold exchange first in 84.8% of cases (versus 1.6% for BM25 and 0.0% for MiniLM with user-turn indexing) and reaches 76.0% end-to-end Application, compared with at most 16.0% across published retrieval and memory systems. A controlled 20-question comparison disentangles recognizing implicit evidence from ranking it properly: Jev and Qwen3.5-9B both recover 90% of gold exchanges within the top 50 (versus 30% for the BERT-based decision model Laya), yet at rank 1 Jev reaches 85% while Qwen3.5-9B drops to 40%. On a dimension-proportional random sample of 60 MemoryAgentBench questions, Jetriever raises the four-dimension average from 44.3% (ReFind) to 55.4%, at a mean retrieval cost of USD 0.00834 per InMind query and USD 0.03103 per MAB query. Together, these results show that effective utility retrieval hinges on combining broad world knowledge, calibrated ranking probabilities, and affordable full-pool evaluation.

cs.CL

版本更新 · v2arXiv:2609.06490

OracleZoom: Reference-Constrained Recursive Image Super-Resolution Beyond Ground Truth

Shubhashis Roy Dipta, Nobin Sarwar, Md Rizwan Parvez

原文摘要

Recursive image Super-Resolution (SR) reaches extreme magnification by feeding each prediction back into the same model, like zooming into an image again and again. Real ground truth for deep zooms is impractical to capture (a full $256\times$ target would be a 52 GB image), so predictions at deeper zooms have no supervision. Yet the last ground-truth image still contains the region that every deeper zoom magnifies. OracleZoom turns this region into a training signal. Inspired by on-policy self-distillation, the model trains on its own prediction for one recursive step ($4\times \rightarrow 16\times$), and a cross-scale loss projects each $16\times$ prediction back to the last ground truth at $4\times$. This projection can verify only coarse content, so a no-reference quality objective guides the remaining fine detail, and a latent prior stops this objective from adding hallucinated artifacts. On seven datasets, OracleZoom achieves the highest CLIPIQA at every zoom depth (mean 0.713 vs. 0.621 for Chain-of-Zoom), with the largest gains at the deepest zooms, even though training stops at $16\times$ and optimizes a different quality metric. Pooled over two ground-truth-anchored judges, this added detail comes at the same hallucination rate as Chain-of-Zoom at $16\times$ and a lower rate at $4\times$, and removing the cross-scale loss raises hallucination at $16\times$.

cs.CV · cs.AI · cs.CL · cs.LG

版本更新 · v2arXiv:2611.18627

FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agentic Language Models

Jian Hu, Yifan Zhang, Hao Zhang, Binfeng Xu, Shaokun Zhang, Hongqing Peng, Zhiding Yu, Pavlo Molchanov 等 10 位作者

原文摘要

Asynchronous reinforcement learning suits long-horizon agents with irregular rollout times, but group-relative updates require repeated rollouts of each prompt. However, group-based sampling limits prompt coverage at a fixed rollout budget and introduces group synchronization barriers. One rollout per prompt avoids these constraints but requires learning from stale, variable-length trajectories without a sibling baseline. We present FlashREINFORCE, to our knowledge the first critic-free single-rollout framework that scales to long asynchronous training and large models. It combines One-Batch REINFORCE, a Sequence Trust Region, and Sample-Mean Optimization, which provide prompt coverage and batch centering while controlling policy drift and length-amplified negative updates. Each component is easy to implement, and the result is their combination: a batch-mean baseline is stable under sample-mean loss, and each common alternative degrades or collapses training. FlashREINFORCE remains stable through 6,000 updates on DeepSeek-R1-Distill-Qwen-1.5B at policy lag four, and through 1,450 on the 30B MoE Qwen3-30B-A3B at lag eight. Under identical loss, data, and lag it matches a strengthened group-relative baseline within 0.5 points, and with half the rollouts exceeds published synchronous baselines on Qwen2.5-Math-1.5B (38.0 versus 36.3 five-benchmark mean). On Qwen2.5-7B-Instruct it maintains tool use where GRPO stops calling tools and achieves 98.3%/96.5% seen/unseen ALFWorld success.

cs.CL · cs.LG

版本更新 · v4arXiv:2607.21653

MOLT: A Hugging Face-Native, Minimal, and Scalable Framework for Agentic Reinforcement Learning

Jian Hu, Huiying Li, Hao Zhang, Binfeng Xu, Yifan Zhang, Shaokun Zhang, Hemil Desai, Michael Demoret 等 11 位作者

原文摘要

Agentic reinforcement learning (RL) research iterates quickly on agents, rewards, and learning algorithms, while policies have grown to hundreds of billions of parameters and beyond. Existing open frameworks force a choice: those that are easy to modify do not scale to the largest models, and those that scale rely on Megatron, which needs dedicated bridge code for every model family and several times more framework code. We present MOLT, an open-source framework built for research that is Hugging Face-native, minimal, and scalable. Its code is easy to modify: the entire RL path is approximately 9.2K lines of Python with one training backend and one rollout engine, so a researcher or an AI coding agent can read it in full and edit the code that actually runs. A Hugging Face checkpoint is loaded directly into a PyTorch FSDP2 trainer with tensor, expert, and context parallelism, so a new model family needs no bridge, and the same code scales from a 4B model to a 753B one. Existing OpenAI- and Anthropic-SDK agents are trained by redirecting their client, with automatic handling of agents that rewrite their own context; training is fully asynchronous; and trajectory data stay in a distributed object store. With MOLT, a 30B mixture-of-experts tool agent trains stably for 1,464 asynchronous updates, a visual computer-use agent improves on OSWorld, asynchronous training is 2.40 times faster than synchronous training, and pipeline throughput reaches 92% of a Megatron-based stack. Its trainer, NeMo AutoModel, trains models of one to 2.8 trillion parameters.

cs.LG · cs.CL · cs.DC

版本更新 · v5arXiv:2410.23066

Don't Pay Attention, PLANT It: Pretraining Attention via Learning-to-Rank

Debjyoti Saha Roy, Byron C. Wallace, Javed A. Aslam

原文摘要

State-of-the-art Extreme Multi-Label Text Classification models rely on multi-label attention to focus on key tokens in input text, but learning good attention weights is challenging. We introduce PLANT (Pretrained and Leveraged AtteNTion), a plug-and-play strategy for initializing the label-attention module that operates on top of an LLM's output representations. PLANT works by planting label-specific attention using a Learning-to-Rank model trained on mutual information gain, before end-to-end fine-tuning. This architecture-agnostic approach integrates with LLM backbones (e.g., Mistral-7B, Llama-3.1-8B, DeepSeek-V3, and Phi-3-mini). PLANT outperforms SOTA methods across tasks like ICD coding, legal topic classification, and content recommendation, and improves fine-tuned LLMs in controlled comparisons outside the medical domain. Gains are largest on rare labels, although a training-free LLM extraction baseline remains competitive on rarely seen labels. Ablation studies confirm that attention initialization is a key driver of these gains. We make our code available at https://github.com/deb-official/plant

cs.CL · cs.LG

新论文 · v1arXiv:2611.23835

NarrowSpec: Draft Hidden Width Need Not Match the Target in Speculative Decoding

Alagappan Valliappan, Ankit More, Dheevatsa Mudigere, Jatin Mitra

原文摘要

Speculative decoding is widely used to speed up generation in large language models (LLMs), with a small draft model proposing tokens that one forward pass of the target model verifies. Draft models that read the target's hidden states, such as DFlash and EAGLE-3, typically inherit the target's hidden width. We show that the draft's whole residual stream can run at half or a quarter of that width while largely preserving acceptance. For DFlash, which borrows the target's vocabulary projection, the trained output adapter folds into it, giving one narrower projection with logits unchanged up to floating-point rounding. We test two drafter families, DFlash and EAGLE-3, on two targets, Qwen3-8B (NVFP4) and GPT-OSS-20B (MXFP4), in SGLang on a B200 and a DGX Spark (GB10). At batch size 1 and temperature 0, narrowing the draft to half or a quarter width speeds up end-to-end decoding by 1.06-1.12x on B200 and 1.13-1.44x on GB10. The speedup holds at larger batch sizes. From batch 32 to the largest batch each GPU fits, narrowing reaches up to 1.21x on both GPUs, and no narrow draft is slower than full width.

cs.CL · cs.LG · cs.PF

新论文 · v1arXiv:2611.23832

A System 1 Decision Model for Medical Multiple-Choice Questions Built on Frozen Language Models

Pavan Kumar Mantha

原文摘要

A System 1 model answers a multiple-choice question in a single pass. A frozen language model turns the question and each option into vectors, two small, trained heads place them in a shared space, and the closest option wins. The design has been shown on agent decisions, but we are not aware of a study of how far it goes on exam questions that need medical knowledge, where the right option is often no more similar to the question than the wrong ones. We rebuilt the recipe from scratch for MedMCQA and trained only two projection heads and one temperature on top of frozen Qwen3.5-0.8B, Granite-4.2-3B and Qwen3-8B, using a contrastive loss with in-batch and hard negatives and a mask for repeated answers. With plain last-token vectors, both small encoders answer about one question in three correctly (33.1% and 33.0% on 3,842 held-out questions, where chance is 25.0%). Writing each text inside a one-word-summary prompt and lowering the learning rate of the heads raised this to 41.7% and 42.5% for the two small encoders and to 46.0% +-0.2 (three seeds) for Qwen3-8B. Shuffling the questions against the options returns accuracy to chance, so the heads do use the question. Flan-PaLM and Med-PaLM 2 report 57.6% and 72.3% on MedMCQA under a different protocol, so ours is a cheap, fast baseline and not a competitive system. We also say what our evidence cannot show: validation gains were about 1.7 to 2 times the test gains, and we could not separate the effect of reading several encoder layers from the other changes.

cs.AI · cs.CL