PyTorchKR Papers

업데이트된 논문

총 23편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2607.23132v2cs.CV

DispatchRAG: 교통사고 영상으로부터 실세계 프로토콜에 기반한 긴급 출동 결정의 그라운딩

DispatchRAG: Grounding Emergency Dispatch Decisions in Real-World Protocols from Traffic Accident Video

Assessing the severity of a traffic accident scenario is important to decide which emergency service to dispatch. Missing an ambulance dispatch on a pedestrian accident is a fatal issue that can lead to death. Recently, Vision-Language Models (VLMs) have been a promising tool for accident reasoning, yet many VLMs are not grounded in real-life accident response protocols, making them not usable in accident severity assessment off-the-shelf. We introduced DispatchRAG, a decision-support framework grounded in real-life Japanese traffic-accident response protocols, designed to assist emergency dispatchers by using VLMs to generate an appropriate emergency response recommendation during an emergency scenario. Utilizing a RAG-based retrieval mechanism to retrieve the most relevant accident protocol and an LLM-powered reasoner to suggest the most proper response. To support evaluation, we introduce Accident Dispatch Dataset, a comprehensive dataset of accident assessment and emergency response according to Japanese accident response protocols adapted from the MM-AU dataset. We validate our framework on the Accident Dispatch Dataset, showing strong performance across various accident scenarios compared to the baseline VLM, pointing toward a decision-support tool that can help autonomous vehicles and human dispatchers respond faster to both their own and nearby accidents.

arXiv:2607.12818v3cs.CV

Breaking D\'ej\`a Vu: Independent Auditing of Visual Place Recognition through Vision-Language Reasoning

시각적 장소 인식(VPR)은 동시 위치추정 및 지도작성(SLAM)을 위한 루프 클로저 탐지와 같이 로봇 공학 응용에서 정확한 위치 추정과 장기 자율 주행을 가능하게 하는 핵심 기술이다. 그러나 실제 환경에서의 VPR 배포는 정밀도와 재현율의 균형을 맞추는 이미지 매칭 임계값 선택에 의존한다. 이러한 임계값은 일반적으로 라벨이 붙은 검증 데이터로 조정된 뒤 배포 시 고정되므로, 정답(ground truth)을 사용할 수 없는 환경 변화 하에서는 신뢰할 수 없다. 이는 거짓 루프 클로저를 수용하면 추정 궤적과 지도가 손상될 수 있는 안전이 중요한 로봇 공학에서 특히 문제가 된다. 본 연구에서는 쿼리 이미지와 후보 이미지에 대해 공동으로 추론하여 검색된 매칭을 평가하기 위해 비전-언어 모델(VLM)을 활용하는, 검색 후 독립 검증 프레임워크인 Visual Place Recognition Auditing을 제안한다. 기존 검증 방법과 달리, 본 접근법은 아키텍처별 신뢰도 측정, 데이터셋 의존적 임계값, 또는 배포 환경에 대한 사전 지식 없이 인스턴스 수준 검증을 수행한다. 우리는 5가지 최신 VPR 방법과 4개의 VLM을 사용하여 6개 벤치마크 데이터셋에서 방법을 평가했다. 결과는 VLM 기반 감사가 최신 방법 대비 평균 recall@1을 13.6% 향상시키면서 거짓 수용률을 12%로 낮추고, 정밀도 95% 이상과 커버리지 75% 이상을 유지함을 보여준다.

arXiv:2601.00821v5cs.AI

구조보다 충실도: 장기 대화 LLM 메모리에서 원문 청크가 손실성 아티팩트 추출을 능가한다

Fidelity Before Structure: Verbatim Chunks Beat Lossy Artifact Extraction in Long-Conversation LLM Memory

A growing class of conversational-memory systems compresses dialogue history into structured artifacts (extracted facts, decisions, or events). The premise is that distilled structure retrieves better than raw text. We test this premise with a controlled ablation. Within one fixed retrieval--rerank--reasoning pipeline, we swap only the stored representation (LLM-extracted typed artifacts versus verbatim conversation chunks) and hold the model, retriever, reranker, and judge constant. Verbatim chunks win by 15.9 points on LoCoMo (43.9% vs. 28.0%) and by 22.0 points on LongMemEval-S (67.4% vs. 45.4%). A semantic graph does not recover the gap, and six confound controls reproduce the effect. The mechanism is lossy distillation rather than structure per se. Accuracy tracks how much source text survives in the store, and the extracted-artifact pipeline does not beat naive RAG in overall accuracy (chunks do abstain worse; see Limitations). For the extraction designs we test, structured memory should augment verbatim text rather than replace it. Adding artifacts alongside chunks costs no detected accuracy, while substituting them forfeits the gap. Code and data are available at https://github.com/tao-hpu/cog-canvas.

arXiv:2303.16880v4cond-mat.dis-nn

Storage and Learning phase transitions in the Random-Features Hopfield Model

The Hopfield model is a paradigmatic model of neural networks that has been analyzed for many decades in the statistical physics, neuroscience, and machine learning communities. Inspired by the manifold hypothesis in machine learning, we propose and investigate a generalization of the standard setting that we name Random-Features Hopfield Model. Here $P$ binary patterns of length $N$ are generated by applying to Gaussian vectors sampled in a latent space of dimension $D$ a random projection followed by a non-linearity. Using the replica method from statistical physics, we derive the phase diagram of the model in the limit $P,N,D\to\infty$ with fixed ratios $\alpha=P/N$ and $\alpha_D=D/N$. Besides the usual retrieval phase, where the patterns can be dynamically recovered from some initial corruption, we uncover a new phase where the features characterizing the projection can be recovered instead. We call this phenomena the learning phase transition, as the features are not explicitly given to the model but rather are inferred from the patterns in an unsupervised fashion.

arXiv:2410.13013v2cs.CL공식 코드 · Papers with Code

LEGAL-UQA: Testing Constitutional Question-Answering Systems for Urdu Speakers with a Parallel English-Urdu Benchmark

Legal information assistants could help Urdu speakers understand their constitutional rights, but building reliable systems requires test data and evaluation methods suited to the language. We present LEGAL-UQA, a parallel English-Urdu benchmark of 619 question-answer pairs linked to passages from Pakistan's Constitution. Each example includes the question, reference answer, and source passage in both languages. We evaluate two components separately: answering Urdu questions when the source passage is supplied, and retrieving the passage from which each question was generated. We also compare JEV, a decision model used as an automated judge, with human grades. The findings show that model rankings depend on how answer similarity is measured, while a simple word-matching method retrieves source passages more effectively than the embedding methods tested. At the evaluated cutoff, JEV agrees with human grades less often than a baseline that gives every answer full credit. An audit also identifies cases where generated references omit information supported by the source. Together, these findings highlight the need to test retrieval and answering separately, include simple retrieval baselines, and validate automated judges and reference answers with human review when engineering legal QA software for Urdu speakers.

arXiv:2607.02338v3cs.DB

Graph-Based Nearest-Neighbor Search with Recall Guarantees Using Graph Spanners

Graph-based indexes such as Hierarchical Navigable Small World (HNSW) graphs make nearest-neighbor retrieval efficient, but a high average recall can hide queries whose results fall below an application's quality requirement. We introduce Certify-then-Rectify (CTR), which returns an initial graph-search result when a calibrated rule accepts it and searches for missing neighbors otherwise. The recovery search reuses the existing graph and the distances already evaluated by the initial search. To determine how far this search should proceed, we relate metric distances to graph-path lengths through graph stretch. We develop Stretch-Bounded Expansion from Query (SBE-Q), together with Metric Bound Verification (MBV) to prune unnecessary work, and estimate its budget from paths to the fraction of true neighbors needed for the requested recall. Finite-sample tolerance bounds support this recall-aligned estimate; paths recorded during proposal recovery offer a conservative alternative to separate shortest-path computations. Independent tests certify initial acceptance and recovery outcomes, giving an end-to-end population recall guarantee even when the estimated stretch does not cover every relevant path. If recovery cannot be certified, rejected queries use exact scan. Experiments on HNSW compare five stretch estimators, assess recovery efficiency, and evaluate compliance and throughput on databases of up to 100 million vectors. CTR is not specific to HNSW, and we also evaluate it on another graph index.

arXiv:2604.22294v3cs.CL

SLIDERS: Evidence Synthesis with Autonomous Database Reconciliation for Systematic Reviews

Systematic reviews requires comprehensive evidence synthesis from large document corpora to answer targeted research questions and are foundational in finance, social sciences, and other technical fields. Manual construction of evidence tables is labor-intensive, and recent LLM-based assistants relying on embedding or keyword search often fail to meet the coverage standards of systematic reviews. We introduce SLIDERS, a novel LLM-based methodology for automatically assembling evidence tables tailored to research questions for systematic reviews. SLIDERS extracts full-text excerpts that serve as direct evidence or as provenance for structured data. Core to SLIDERS is an autonomous evidence reconciliation coding agent that iteratively writes SQL to analyze and reconcile extracted evidence, bringing together information fragmented across documents, resolving inconsistencies and removing duplicates across records. We evaluate SLIDERS on BrowseComp+, achieving 90% accuracy, and show that it outperforms all baselines on additional benchmarks within the LLM's context window. To assess scalability, we introduce two new evidence-synthesis benchmarks, WikiCeleb100 (3.9M tokens) and FinQ100 (36M tokens), which include follow-up questions beyond the original research query. Although SLIDERS synthesizes evidence knowing only the original question, it answers 78.9% and 55.2% of all questions correctly on WikiCeleb100 and FinQ100, respectively

arXiv:2607.22639v2cs.AI

TRACE: 엔터프라이즈 LLM에서 지식을 보존하는 파라미터릭 도구 검색을 위한 비즈니스 규칙 기반 추론 커리큘럼

TRACE: Business Rule-Grounded Reasoning Curriculum for Knowledge-Preserving Parametric Tool Retrieval in Enterprise LLMs

파라미터릭 검색(parametric retrieval)은 각 API에 고유한 가상 토큰을 할당하고, 제약 빔 탐색(constrained beam search)을 통해 모델이 이를 생성하도록 학습함으로써 LLM이 도구를 암묵적으로 검색할 수 있게 한다. ToolSense는 이 방식이 두 가지 치명적인 단점을 갖는다고 보여준다: 학습(training) 중 파라미터릭 도구 지식을 파괴하고, 빔 탐색 디코딩이 실시간 배포에 너무 느리다는 것이다. 우리는 이러한 분리(dissociation)를 해소하는 2단계 커리큘럼인 TRACE(Tool Retrieval via Augmented Chain-of-thought and Enterprise rules)를 제안한다. 1단계는 ToolSense의 다중 형식 암기 SFT를 재사용하여 LoRA로 도구 지식을 시드(seed)한다. 2단계가 핵심 기여이다: 모델은 ToolSense의 RRB 쌍과 도메인 전문가가 큐레이션한 비즈니스 규칙을 대상으로 합성된 쿼리라는 두 가지 데이터 소스를 사용하며, 둘 다 추론 trace로 증강된 데이터로, 도구 토큰의 JSON 목록을 생성하기 전에 thinking trace를 출력하도록 학습된다. 이 학습 목표는 1단계 MCQ 및 QA 프로빙 정확도를 보존하면서 프로덕션 지연 시간에서 단일 빔 greedy 디코딩을 가능하게 한다. 두 엔터프라이즈 제품군에 걸친 8,300개 이상의 도구를 포함한 통합 엔터프라이즈 카탈로그에서 평가한 결과, 2단계 TRACE 학습은 도구 이해력을 보존할 뿐 아니라 향상시킨다: MCQ 정확도는 1단계 대비 +3.2 pp, QA 프로빙은 +9 pp 향상한다. 검색(retrieval)에서는 TRACE가 Domain A에서 ~86% 재현율, Domain B에서 ~60% 재현율을 달성한다—임베딩(embedding) 기준선(baseline) 성능 ~27% 및 ~52%와 비교—둘 다 단일 빔 greedy 디코딩으로, 프로덕션 지연 시간에서 직접 배포 가능하다.

arXiv:2609.27710v3cs.CV

FFM-CP: 퓨샷 계산병리학을 위한 비전-언어 기반 모델의 교차 백본 융합

FFM-CP: Cross-Backbone Fusion of Vision-Language Foundation Models for Few-Shot Computational Pathology

병리학 비전-언어 기반 모델은 질환과 과제에 따라 성능이 달라지며, 어떤 단일 모델도 일관되게 최고 성능을 내지 못한다. 전문가 병리 주석의 높은 비용은 과제별 적응에 사용할 수 있는 라벨 데이터를 제한하기도 한다. 상호 보완적인 사전학습 표현을 결합하는 것은 이러한 한계에 대한 잠재적 접근이지만, 소수의 라벨 예시로부터 효과적인 융합을 학습하는 것은 여전히 어렵다. 본 연구는 퓨샷 학습 설정에서 여러 병리학 비전-언어 모델을 결합하는 프레임워크인 Few-shot Fusion Foundation Models of Computational Pathology(FFM-CP)를 제안한다. 이 프레임워크는 먼저 대응하는 서포트 이미지로부터 추정한 폐형식(closed-form) 직교 프로크루스테스(Orthogonal Procrustes) 변환을 사용하여 이질적인 표현을 정렬한다. 이 정렬은 추가 정렬 네트워크를 학습하지 않고도 모델 내부 특징 기하를 보존한다. 정렬된 공간 안에서 통합 그래프는 서포트 이미지 특징과 시각·텍스트 클래스 프로토타입을 공동으로 정제함으로써 백본 간 정보 교환을 가능하게 한다. 이렇게 정제된 표현은 의미적 클래스 지식과 클래스 내 시각적 변이를 각각 포착하는 상호 보완적인 텍스트-프로토타입 분기와 케이스 검색(case-retrieval) 분기를 지원한다. 각 분기는 모든 순서화된 백본 쌍의 예측을 결합하는 방법을 학습하여, 한 모델로 인코딩된 쿼리가 다른 모델이 표현한 증거를 활용할 수 있게 한다. 우리는 6개의 조직병리학 데이터셋에서 클래스당 4, 8, 16샷 조건으로 3가지 백본 조합을 평가했다. FFM-CP는 54번 비교 중 50번에서 각 융합 집합의 개별 적응 멤버 중 가장 강한 모델보다 더 높은 평균 macro-F1을 달성했다. 이러한 결과는 주석이 제한될 때 상호 보완적인 사전학습 표현의 결합이 조직병리학 분류를 개선할 수 있음을 시사한다.

arXiv:2607.13298v2cs.CV

FOLIO: 스트리밍 비디오 이해를 위한 집중 의미 기억

FOLIO: Focused Semantic Memory for Streaming Video Understanding

온라인 스트리밍 비디오 이해에서는 비디오 스트림이 계속 도착하고 질의가 언제든지 발생할 수 있다. 스트리밍 프레임이 무한히 증가하기 때문에, 시스템은 미래 프레임과 미래 질의가 알려지지 않은 상태에서 관측된 비디오 접두 문맥의 정보를 지속적으로 압축하고 보존해야 한다. 핵심 과제는 어떤 정보를 보존하고 유지된 이력을 어떻게 구성할지 결정하는 것이다. 스트림에 따라 이력이 커질수록 메모리 비용이 증가하고 많은 중복 시각적 세부 사항이 보존되는 반면, 이후 질의는 종종 특정 개체, 행동 및 그들의 시간적 변화에 의존한다. 이 과제를 해결하기 위해, 우리는 스트림의 중요한 부분을 더 높은 상세도로 기록하면서 주변 컨텍스트는 압축적으로 유지하는 학습 불필요 집중 의미 기억 시스템인 FOLIO를 제안한다. 스트림이 도착함에 따라 FOLIO는 동적 초점 상태에 의해 안내되어 세그먼트 수준에서 메모리를 갱신하며, 단기 시각 버퍼와 관측된 개체를 중심으로 구성되고 시각적 증거 캐시에 연결된 장기 의미 기억을 결합한다. 질의 시점에는 경량 하이브리드 검색이 구조화된 메모리에 대한 직접 매칭과 의미적 질의 확장을 결합한다. FOLIO는 최첨단 성능을 달성하며, Qwen3-VL-8B로 OVO-Bench에서 인식/역방향 정확도 82.0/69.1을 기록하고 StreamingBench에서 전체 정확도 74.5를 달성하는 동시에, 초점 개체에 대해서만 상세 기록을 예약하고 주변 컨텍스트를 압축적으로 저장함으로써 스트리밍 메모리 유지 비용을 크게 줄인다.

arXiv:2607.11976v5cs.LG

LiteTopK: 장문맥 희소 어텐션용 통합 Indexer-TopK 커널을 위한 차원의 저주 활용

LiteTopK: Exploiting the Curse of Dimensionality for a Fused Indexer-TopK Kernel in Long-Context Sparse Attention

Indexer-TopK, which computes scores and selects the top-$k$ candidates, is a key primitive in long-context sparse attention for large language models, vector retrieval for recommendation systems, and vector databases. However, existing GPU-based Indexer-TopK kernels like DeepSeek Sparse Attention (DSA) typically separate scoring from selection and materialize the full score matrix in high-bandwidth memory. This strategy is inefficient for large score matrices, such as those encountered in long-context sparse attention, due to excessive global-memory traffic, costly synchronization, and prohibitive memory overhead. In this study, motivated by the curse of dimensionality, we observe that sparse attention scores exhibit a score concentration phenomenon, with scores falling within a narrow range. Based on this observation, we propose LiteTopK, a sample-filter-select framework that estimates query-data score ranges from a small subset of the data. It uses these estimates to filter scores as they are computed, materializes only the candidates that pass the filter, and selects from the surviving candidates. In particular, we develop two implementations of LiteTopK. The first uses a conservative threshold estimate, while the second pairs a predicted threshold with a fallback strategy. Building on LiteTopK, we further propose LiteDSA, which reuses overlapping key/value entries across neighboring queries to accelerate attention computation. We integrate LiteTopK into vLLM, SGLang, and Megatron-LM. Our evaluation on eight B200 GPUs demonstrates prefill speedups of up to 1.49$\times$ across GLM-5.2, DeepSeek-V4-Flash, and Hy-4-Preview at a context length of 1M tokens, without compromising model capabilities.

arXiv:2604.20133v5cs.AI

EvoAgent: An Evolvable Agent Framework with Skill Learning and Multi-Agent Delegation

Large language model (LLM) agents increasingly rely on reusable skills, but skills are usually written by hand, and letting an agent change them on its own calls for explicit user control. We report on EvoAgent, an evolvable agent framework built for a foreign trade assistant, and share the engineering choices behind it. We organize the agent as a harness of four components (reasoning, context engineering, tools and state) and confine self-improvement to the user's persistent state: profile, memory and skills. Skills are multi-file packages with lazily loaded references, matched in three stages (keyword, embedding, LLM). The main agent can delegate to sub-agents with their own context, and long histories are compressed while an index of recoverable resources is kept. An offline loop reviews completed sessions to update the profile and memory and to propose skills; in the open-source release a skill changes only after the user confirms the proposal. On 172 foreign trade test cases scored by an LLM judge on five dimensions, GPT-5.2 with EvoAgent scored about 28% higher than GPT-5.2 called directly with only the final query. This is a single run judged by the model under test, so it measures the harness as a whole rather than any one mechanism, and swapping in GPT-4.1 or Qwen3.5-35B-A3B lowered the scores relative to calling those models directly. Several mechanisms described here were added after the version we evaluated and are not evaluated. The evaluation data is proprietary, so the results are for reference only. An open-source release of the code, which is not the evaluated business version, is available at https://github.com/Focus-AI-Center/Mentarc-EvoAgent

arXiv:2505.16409v2cs.AI

Retrieval as Exploration: Retriever-Free Retrieval via Corpus-Traversing MCTS

Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in multi-step reasoning and calling search tools at appropriate steps. However, existing retrieval-augmented reasoning approaches still rely on separate retrieval models, limiting LRMs to deciding when to retrieve and how to formulate queries rather than leveraging their pretrained knowledge and reasoning capabilities for retrieval itself. This structural separation introduces a representation bottleneck, as retrieval relies on separately learned representations that may obscure fine-grained distinctions needed for reasoning. Moreover, these approaches commonly employ fixed top-k retrieval, which can introduce superfluous documents and increase retrieval noise and context overhead. To address this, we reformulate retrieval from query-document relevance matching into the exploration of answer-containing paths within the corpus, and propose FREESON (Retriever-FREE Retrieval-Augmented ReaSONing), a framework that enables LRMs to directly navigate the original corpus text to locate information that satisfies their specific needs, serving as both the reasoner and the retriever. Specifically, we introduce CT-MCTS (Corpus-Traversing Monte Carlo Tree Search), a retrieval-specialized variant of Monte Carlo Tree Search (MCTS). CT-MCTS iteratively expands and evaluates exploration paths, starting from multiple promising entry points across the corpus. This allows LRMs to selectively reference specific segments identified during traversal, instead of fetching a fixed top-k set. Across nine benchmarks, FREESON improves average EM and F1 by 15 points over four retriever-based baselines on five general-domain QA tasks, matching the strongest baseline. It also gains 12 points on fact-checking and 13 points on medical QA under zero-shot transfer, performing on par with the much larger MedGemma-27B.

arXiv:2610.02945v2cs.AI

Continual Graph Memory and Community Infrastructure for AI-Assisted Mathematical Discovery

Mathematical discovery requires preserving partial ideas and the evidence supporting their reuse. We present Ansatz, a research harness whose continual graph memory separates model-accepted claims and dependencies from curated exploration, and Ansatze, a community platform for inspecting, guiding and checking this work. On First Proof, Ansatz recorded nine local verifier closures in each of two ten-problem batches; a partial result in the Second Batch also counted as a pass under the study's criterion. These model-assessed outcomes include continued searches and differ from expert validation. Component studies and re-review expose disagreements and their dependency reach. Open-problem investigations produced candidate results, with five Lean formalizations reported. The platform connects problem discovery, research records, manuscript analysis and formal checking while preserving distinct evidence states. Together, these systems support cumulative human-AI research through persistent investigation and inspectable contributions.

arXiv:2501.08686v2cs.DB공식 코드 · Papers with Code

Knowledge Graph-based Retrieval-Augmented Generation for Schema Matching

Traditional similarity-based schema matching methods fail to resolve semantic ambiguities and conflicts in complex, domain-specific mappings due to a lack of commonsense and domain-specific knowledge. In parallel, large language models (LLMs) without sufficient background knowledge tend to hallucinate, further complicating schema matching. To overcome these limitations, we introduce KG-RAG4SM, a Knowledge Graph-based Retrieval-Augmented Generation for Schema Matching. This method integrates novel retrieval mechanisms, including LLM-based, vector-based, graph traversal-based, and hybrid techniques with advanced reranking schemes to extract the most relevant subgraphs from large external KGs. Our approach achieves superior results without requiring re-training, as evidenced by improvements of 39.79% in precision and 30.56% in F1 score over LLM-based state-of-the-art (SOTA) methods (e.g., Jellyfish-8B) on the MIMIC dataset, and enhancements of 69.20% in precision and 21.99% in F1 score over pre-trained language model (PLM)-based SOTA methods (e.g., SMAT) on the Synthea dataset. Additionally, KG-RAG4SM scales effectively across KGs from commonsense KGs to domain-specific ones, while its instruction-tuning knowledge exclusion component ensures truthful performance in end-to-end schema matching tasks despite the suboptimal relevance of KG context. Our empirical analysis and case studies from real-world scenarios show that the hallucination problem in LLMs for schema matching is well mitigated by our solution.

arXiv:2605.10832v3cs.CL

비전 네이티브 멀티모달 딥 서치 에이전트를 위한 온-정책 데이터 진화

Towards On-Policy Data Evolution for Visual-Native Multimodal Deep Search Agents

Multimodal deep search requires an agent to solve open-world problems by chaining search, tool use, and visual reasoning over evolving textual and visual context. Two bottlenecks limit current systems. First, existing tool-use harnesses treat images returned by search, browsing, or transformation as transient outputs, so intermediate visual evidence cannot be re-consumed by later tools. Second, training data is usually built by fixed curation recipes that cannot track the target agent's evolving capability. To address these challenges, we first introduce a visual-native agent harness centered on an image bank reference protocol, which registers every tool-returned image as an addressable <image:N> reference and makes intermediate visual evidence reusable by later tools. On top of this harness, On-policy Data Evolution (ODE) runs a closed-loop data generator that refines itself across rounds from rollouts of the policy being trained. This per-round refinement makes each round's data target what the current policy still needs to learn. The same framework supports both diverse supervised fine-tuning data and policy-aware reinforcement learning data curation, covering the full training lifecycle of the target agent. Across 8 multimodal deep search benchmarks, ODE improves the Qwen3-VL-8B agent from 24.9% to 39.0% on average. At 30B, ODE raises the average score from 30.6% to 41.5%. Further analyses validate the effectiveness of image bank reuse, especially on complex tasks requiring iterative visual refinement, while rollout-feedback evolution yields more grounded SFT traces and better policy-matched RL tasks than static synthesis.

arXiv:2606.24428v2cs.CL

Beyond Self-Reflection: An Execute-Distill-Verify Framework for Agentic Experience Learning

Experience learning lets large language model agents reuse lessons through external memory without updating model parameters. However, reflection can endorse execution errors and store them as future guidance, a failure mode we call the Self-Confirmation Trap. We propose EDV, an Execute-Distill-Verify framework for offline experience construction. Heterogeneous agents independently execute each task; a third-party distiller compares their trajectories and environmental feedback to extract candidate lessons; and the executors separately review each candidate, retaining only unanimously approved lessons. Evaluation uses a single solver retrieving from the frozen memory bank. Experiments on $\tau^2$-bench, Mind2Web, and MMTB show improved revisited-task performance, with additional evidence of within-domain transfer to held-out tasks on $\tau^2$-bench. On an annotated Retail subset, EDV reduces the Self-Confirmation Trap Rate from 11.0% with RB-MaTTS to 5.0%, indicating less retention of erroneous guidance.

arXiv:2606.21633v3cs.LG

HERALD: High-Throughput Block Diffusion LLM Inference via CPU-GPU Cooperative KV Cache Retrieval

Block diffusion LLMs (block dLLMs) improve decode throughput by generating tokens in parallel, and are emerging as a next-generation alternative to autoregressive LLMs. At long contexts, however, the KV cache remains a major throughput bottleneck in both memory bandwidth and capacity. For autoregressive LLMs, dynamic sparse attention is often combined with KV offloading to alleviate both bottlenecks, keeping the full KV cache in CPU DRAM and fetching only a critical subset to the GPU. However, KV subset selection and retrieval over bandwidth-limited PCIe incur substantial offloading overheads that often stall the GPU. We present HERALD, the first dynamic sparse attention method for block dLLMs that supports KV offloading. HERALD exploits a property of block dLLMs: a critical KV subset selected once per block is reused across all of its denoising steps, so the next block's selection can be hidden behind the current block's denoising. HERALD realizes this with a two-stage selection algorithm co-designed with its execution system. While the GPU denoises the current block, the CPU speculatively searches the full KV cache for the next block with a single query merged from the current block's B queries, cutting the selection compute by Bx; the GPU then refines the candidates into the final top-K subset with the next block's real queries. At K=1024, HERALD achieves near-lossless accuracy relative to Dense on LongBench and reasoning tasks across four block dLLMs, and improves decode throughput by up to 3.70x over Dense GPU-only inference, 2.59x over MAGE, a state-of-the-art block-dLLM sparse attention method, and 8.38x over MAGE-Offload, its direct KV-offloading extension.

arXiv:2606.19847v2cs.CL

AtomMem: 원자적 사실(Atomic Facts)을 통한 대규모 언어 모델(LLM) 에이전트용 간단하고 효과적인 메모리 시스템 구축

AtomMem: Building Simple and Effective Memory System for LLM Agents via Atomic Facts

대규모 언어 모델(LLM)은 강력한 추론 및 생성 능력을 보여주지만, 고정된 컨텍스트 윈도우(context window)로 인해 다중 세션 상호작용 간의 장기 정보 축적 및 재사용이 제한됩니다. 기존 메모리 증강 시스템은 비효율적인 메모리 표현이나 불안정한 제약 없는 업데이트에 의존하여, 종종 대략적이고 불안정한 방식으로 메모리를 구성합니다. 이러한 문제를 해결하기 위해, 우리는 가치 밀도가 높은 저장과 안정적인 메모리 진화를 위해 설계된 장기 메모리 시스템인 AtomMem을 제안합니다. AtomMem은 Fact Executor를 도입하여, 장문 상호작용에서 고가치 원자적 사실(atomic facts)을 선택적으로 추출함으로써 매우 효율적인 메모리 표현으로 활용합니다. 이후 AtomMem은 이러한 사실들을 계층적 이벤트 구조와 시간적 프로필로 조직하여, 일관된 에피소드 컨텍스트를 포착하고 시간이 지남에 따라 동적으로 진화하는 사용자 속성을 추적합니다. 검색(retrieval) 과정에서 시스템은 연상 메모리 그래프를 활성화하여 파편화된 메모리를 연결합니다. LoCoMo 벤치마크에 대한 실험 결과, AtomMem은 다양한 추론 작업에서 최첨단(state-of-the-art) 성능을 달성하는 것으로 확인되었으며, 이는 지능형 개인화 에이전트 배포를 위한 확장 가능하고 경제적으로 실현 가능한 솔루션을 제공합니다.

arXiv:2606.11806v2cs.CL

프로덕션 LLM 시스템에서의 외부 경험 서빙: 품질-비용 트레이드오프에 대한 배치 지향 연구

External Experience Serving in Production LLM Systems: A Deployment-Oriented Study of Quality-Cost Trade-offs

프로덕션 대규모 언어 모델(LLM) 시스템은 재사용 가능한 운영 경험을 축적하지만, 실제 배치 문제는 단순히 그러한 경험이 도움이 되는지 여부가 아닙니다. 문제는 현실적인 제약 조건 하에서 서로 다른 서빙 전략들이 품질과 온라인 비용을 어떻게 상충(trade-off)시키는가입니다. 외부 경험을 주입하면 작업 품질을 향상시킬 수 있지만, 프롬프트 부담, 지연 시간(latency), 및 서빙 압력 또한 증가시킵니다. 우리는 외부 경험 서빙을 배치 지향 품질-비용 트레이드오프 문제로 연구합니다. 우리는 실제 프로덕션 모더레이션(moderation) 설정에서 이 문제를 평가하며, 서로 다른 출력-비용 체제(regimes)를 드러내는 보조 대비 작업으로 툴 사용(tool-use)과 GPQA를 사용합니다. 우리는 경험 없는 베이스라인, 무작위 경험 통제군, 전역 프롬프트 주입, 검색 기반 선택적 주입을 비교하고 작업 품질과 서빙 비용을 모두 분석합니다. 결과에 따르면, 경험이 사례 의존적이게 되면 선택적 검색이 무조건적인 전역 주입보다 더 강력한 운영 지점을 제공합니다. 또한 검색 품질이 단순히 Top-$K$를 늘리는 것보다 더 중요하며, 동일한 서빙 정책이 짧은 출력(short-output) 및 디코딩 집중(decode-heavy) 체제에서 상당히 다른 비용-편익 프로필을 나타낼 수 있음을 보여줍니다. 이러한 발견은 외부 경험을 보편적인 추가 기능이 아닌, 선택적이고 비용 인식적인 서빙 결정으로 취급하는 것이 가장 좋음을 시사합니다. 전반적으로, 여기서 연구된 설정에서 외부 경험은 서빙 인터페이스와 작업별 비용 구조가 품질 향상이 온라인 비용을 감당할 만한 가치가 있게 만들 때만 효과가 있습니다.