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

+816新增论文
+642其中 AI 相关
+95已有论文修订
+95新增 AI 导读

arXiv 最近完成扫描:2026-11-21 12:19 · 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 2026-11-20 13:14 — 2026-11-21 13:14(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:2026-11-21 12:19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 4 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-21 12:17(北京时间)
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: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:2611.23907

Evaluating invariance training with certified answer-changing controls in neural graph algorithms

Zhe Dong, Fang Qin

原文摘要

Invariance training aims to make neural algorithmic reasoners robust to answer-preserving input changes and is judged by accuracy and by how rarely such edits change predictions. We ask whether this evaluation detects lost sensitivity to answer-changing edits on minimum spanning trees of synthetic graphs, where both directions can be certified. Attaching a subgraph through one bridge provably preserves every original label. Attaching it through two edges, oracle-selected to change a label, and lowering one weight inside the graph serve as answer-changing controls; a response is a changed prediction on an edge whose label changes. A reproduced graph transformer (GDT) within its published accuracy range changes a provably fixed label on 53.3-65.9% of the 64-vertex graphs it solves exactly after a 64-vertex bridge attachment; removing the connecting edge changes little. With 16-vertex graphs and attachments, three invariance objectives lower GDT violations from 2.91% to 0.52-2.02% of pairs at essentially unchanged accuracy. Training also lowers the response to two-edge attachments from 97.6% to 66.0-79.0%, while the response to in-base weight edits stays near 99.5%. At 64 vertices accuracy also falls, but on 16-vertex bases only the attachment control reveals the loss. A message-passing network loses some attachment response at small penalty weights but gains little invariance. Running the conventionally trained GDT on each biconnected block separately removes all observed violations and keeps accuracy and both responses, so the task does not force the trade-off. Sensitivity checks for trained invariance should include certified answer-changing edits that attach material as the training edits do.

cs.AI

新论文 · v1arXiv:2611.20772

Building Fully On-Device GUI Agents with Mobile NPUs

Zhe Dong, Guanyu Cai, Yuchen Jiang, Jiliang Wang

原文摘要

Mobile GUI agents are increasingly used to automate everyday smartphone tasks, but most existing systems still rely on cloud-hosted models or remote inference, introducing repeated communication latency, privacy exposure, and dependence on network availability. Modern smartphones integrate neural processing units (NPUs) that provide an attractive substrate for fully on-device LLM inference. However, directly deploying conventional GUI-agent pipelines on mobile NPUs is inefficient. GUI-agent inference is fundamentally a long-input, short-output workload, yet conventional agents spend substantial time on autoregressive reasoning and action generation, while mobile NPUs provide much higher throughput for prompt prefill than token-by-token decoding. Meanwhile, dynamically varying GUI-agent contexts conflict with the fixed context profiles required by statically compiled NPU runtimes. We present DroidPrefill, a prefill-centric inference architecture that reshapes the GUI-agent workload around mobile-NPU execution. DroidPrefill combines Prefill-Centric Inference to reduce unnecessary autoregressive decoding with Efficient Context Management to preserve decision-critical information within a fixed compiled context budget. Experiments on commercial smartphones show that, in controlled comparisons with the autoregressive T3A+ baseline, DroidPrefill reduces per-step inference latency by up to 85.0%, while improving task success rate by up to 24.6%.

cs.AR