PyTorchKR Papers

최신 논문

총 568,830편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2608.26495v3cs.CV

Video-FLAIR: Not Whether to Reason, But How

Multimodal queries can require different types of reasoning. Some can be answered via perceptual reasoning, extracting information directly from the visual signal, while others require compositional reasoning that combines observations or deliberative reasoning that evaluates competing hypotheses. However, many existing methods apply a uniform reasoning strategy across queries, leading to unnecessary computation on simple tasks and insufficient reasoning on complex ones. We introduce Video-FLAIR, a training framework that learns to select the appropriate reasoning mode for each query using reinforcement learning. During training, the model generates responses under all three modes for the same prompt, enabling direct comparison. A composite reward compares these responses to favor the most effective one based on correctness, grounding, and cost, while discouraging unsupported or misaligned deliberation. This yields a supervision signal for learning adaptive reasoning without per-query annotations. Video-FLAIR improves accuracy over the Qwen2.5-VL base model by +5.4 on MathVista, +4.8 on Video-Holmes, and +4.8 on Video-MMMU, while reducing average token usage to 95 compared to 417 for always-thinking baselines.

arXiv:2608.26086v4cs.LG

TraceML: What Auto-Research Agents Miss in Long-Horizon ML Development

Auto-research agents now run machine-learning development unattended for hours, revising data pipelines, models, and validation from their own feedback, yet on most competitions they still finish below strong human competitors. Outcome-based benchmarks record this gap but not its cause, because they grade the final submission and discard the development process behind it. We introduce TraceML, which pairs human and agent work on the same competitions under one version-level schema: 4,465 human Kaggle trajectories across 134 competitions, seven of which are also worked by two agent scaffolds, giving 430 paired human and 207 agent trajectories. Every code version carries its score, its timestamp, and labels for the action taken, its intent, the edit size, and the score effect. Read this way, the gap becomes concrete. Experts alternate data work, validation, model changes, and ensembling, and return to approaches they had set aside. Each agent scaffold instead collapses into a narrow loop: Codex spends its steps re-weighting ensembles and tuning submissions, MLEvolve mutates its model in place, and neither pivots at the human rate nor reopens abandoned work. A short planning prompt distilled from human practice moves the behaviors it names toward the human profile and lifts scores, but the effort profile stays agent-shaped: instruction closes only the part of the gap that reduces to instructions. We release the corpus, the schema, and the labeling models, together with an open-source toolkit that turns a run from any command-line agent into a TraceML trajectory and reads it against the human cohorts.

arXiv:2608.25100v4cs.AI

다목적 강화학습을 통한 신뢰 가능·일반화·특이성 문맥 내 지식 편집

Towards Reliable, Generalizable, and Specific In-Context Knowledge Editing via Multi-Objective Reinforcement Learning

대규모 언어 모델(LLM)은 강력하지만, 사전학습이 종료된 뒤 시대에 뒤처지는 정적 파라미터 지식 때문에 한계가 있다. 지식 편집(knowledge editing)은 전체 재학습 없이 목표 사실에 대한 모델 동작을 갱신하여 이 문제를 다룬다. 특히 문맥 내 지식 편집(in-context knowledge editing)은 추가 학습이 필요 없고 블랙박스 LLM에도 바로 적용할 수 있어 주목받고 있다. 최근 강화학습(RL) 기반 접근은 양-질 트레이드오프에 맞게 프롬프트 구성을 조정함으로써 고정 검색 전략보다 나아진다. 그러나 초기 성과에도 불구하고, 신뢰성(reliability), 일반화성(generality), 특이성(specificity)이라는 서로 구분되고 종종 상충하는 목표 아래에서 프롬프트를 구조화된 개체로 모델링하지 못한다. 이전 방법들은 대체로 단일 목표만 최적화하고 프롬프트 구성 과정의 일부에 대해서만 결정을 내려, 서로 다른 목표 간 균형과 데모(demonstration)의 전역적 배치를 모두 놓친다. 본 논문은 문맥 내 지식 편집을 위한 프롬프트 구성을 제약 마르코프 결정 과정(Constrained Markov Decision Process)으로 공식화하는 다목적 RL 알고리즘인 Multi-Objective In-context Knowledge Editing(MO-IKE)를 제안한다. MO-IKE는 지식 편집에서 경쟁하는 목표들을 최적화하도록 동적 검색기를 학습하여, 더 균형 잡히고 전역적으로 일관된 프롬프트 구성을 가능하게 한다. Llama-3.2에서 MO-IKE는 선행 RL 기반 방법 대비 편집 성공률(신뢰성)을 85.0%에서 92.0%로, 패러프레이즈 일관성(일반화성)을 77%에서 79%로 높였으며, 보존율(특이성)은 23.0% 증가시켰다.

arXiv:2608.25097v4cs.AI

PhysElite: LLM은 올림피아드 수준의 물리 문제 해결에서 얼마나 멀리 있는가?

PhysElite: How Far Are LLMs from Solving Olympiad-Level Physics Problems?

Understanding how (multimodal) large language models perform on physics problems requires benchmarks that reflect the difficulty and breadth of expert-level physical reasoning. Existing physics benchmarks remain limited in the following two important ways: (1) short of high-difficulty datasets, and (2) lack of comprehensive coverage of visual forms, knowledge points, and step-by-step solution processes. As a result, model performance on current datasets may not be fully representative of their ability to solve complex physics problems. To address these issues, we present PhysElite, a large-scale bilingual multimodal benchmark for Olympiad-level physics reasoning. PhysElite contains 11,586 Olympiad-tier problems. For each problem, we provide corresponding visual diagrams, step-by-step bilingual Chinese-English solution derivations, and the final answer. We benchmark 18 open-source and closed-source MLLMs, and find that even the strongest model reaches only 33.7% answer accuracy. We additionally conduct step-level process evaluation to diagnose where models fail in the reasoning chain. Our datasets are released at https://github.com/PremiLab-Math/PhysElite.

arXiv:2608.25005v2cs.CL

미완료 상의 역설은 대규모 언어 모델에 반드시 존재하는 것은 아니다: 모델 실패 이전에 발생한 벤치마크 실패

The Imperfective Paradox Is Not Necessarily in Large Language Models: A Benchmark Failure Before a Model Failure

Recent work uses an NLI benchmark of the imperfective paradox to argue that large language models exhibit a Teleological Bias and Calibration Crisis. We identify three validity concerns in benchmark construction, including aspectual reduction, semantic bias, and relation bias. Under a strict NLI standard, 76% of the Group A instances do not explicitly rule out culmination; in native-speaker annotation, 38% of Group A and 29% of Group C examples permit an alternative interpretation. To better isolate the semantic contrast, we construct Lexically Matched Minimal Pairs. Moreover, by formulating event-semantic NLI as a multi-step reasoning problem, we evaluate intermediate semantic judgments and use oracle-guided settings. We find that incorrect NLI entailment does not fully support teleological bias. Prompting often changes final NLI decisions without reliably improving the underlying semantic judgments, offering a possible explanation for the Calibration Crisis.

arXiv:2608.24845v2cs.CV

LAION-BVD: A 13-Million-Hour Open Video Dataset for Multimodal Pre-training

We present LAION-BVD (LAION - Big Video Dataset), a large-scale open video dataset for multimodal learning, which contains 1.3B platform-specific video URLs collected from CommonCrawl. From these, we download 80M videos with a total duration of 13.5 million hours. The dataset is designed for multimodal pre-training across the video, audio, and image modalities. Using content-aware scene detection, we extract clips for which we synthetically generate video and audio captions. Models trained on these data achieve competitive performance on standard video-text and audio-text benchmarks, with consistent improvements as training or model scale increases. Additionally, we explore video frames as an alternative source of image-text data by extracting scene-changing frames. These frames exhibit a visual distribution distinct from standard web image corpora, and models trained on this dataset achieve strong image-text retrieval performance. We release LAION-BVD to the research community. It significantly expands open access to multimodal videos at an unprecedented scale.

arXiv:2608.24046v2cs.AI

Algorithmic Impact Reveals the Hidden Social Choice Structure of Alignment

When an AI algorithm makes decisions that affect more than one person, aligning it becomes a problem of social choice: how should people's divergent preferences about system behavior be reconciled and aggregated into a single coherent model? The standard approach to aligning frontier AI models - reinforcement learning from human feedback - largely sidesteps this question and has poor social choice guarantees. However, it remains unclear what alternative should replace it. We show that, when agent utilities can be expressed linearly in a fixed feature representation, the alignment problem can be reformulated as linear optimization over a convex impact space, which makes it amenable to the standard toolkit of welfare economics and mechanism design. This reformulation clarifies how alignment protocols translate into welfare consequences and, conversely, how a social planner's desired constraints on welfare consequences can be translated back into alignment protocols. We apply this transformation to show that voting-by-issues and random-dictatorship mechanisms are strategyproof and unanimous. Demonstrating the reverse direction, we also apply the impact representation to derive a family of alignment protocols that maximize utilitarian social welfare subject to various social desiderata, such as bounds on individual or group harm. We illustrate the welfare implications of these alignment protocols empirically using real human preferences over kidney allocation, charitable food distribution, LLM responses, and trolley problems.

arXiv:2608.24005v2cs.AI

다중 도메인 지식 추적을 위한 인지 부하 및 지식 전이 통합

Incorporating Cognitive Load and Knowledge Transfer for Multi-Domain Knowledge Tracing

지식 추적(KT)은 학습 이력으로부터 학습자의 동적 지식 상태를 평가하는 것을 목표로 한다. 기존 KT 방법 대부분은 단일 도메인 학습에 초점을 두고 상당한 성공을 거두었지만, 실제 학습 상황에서는 여러 도메인을 동시에 다루는 경우가 많아 다음 두 가지 중요 요인이 등장한다: 1) 시간적·지식적 차원 모두에서 여러 도메인에 걸친 학습을 관리하면서 생기는 인지 부하(cognitive load). 2) 한 도메인의 지식 상태가 도메인 내 및 도메인 간 관련 상태에 영향을 미치는 지식 전이(knowledge transfer). 본 논문에서는 다중 도메인 학습 상황에서 학습자 지식 상태 평가를 개선하기 위해 이러한 요인을 탐구하는 데 초점을 두고, 인지 부하(Load)와 지식 전이(Transfer)를 통합한 다중 도메인 지식 추적(Multi-domain Knowledge Tracing) 방법 LT-MKT를 제안한다. 구체적으로, 고립된 도메인을 연결하기 위해 LT-MKT는 문제와 연관 개념의 텍스트 정보를 통합하여 다중 도메인 계층 그래프(Multi-domain Hierarchical Graph)를 구성하며, 대규모 언어 모델(LLM)의 뛰어난 표현 능력을 활용한다. 이어 시간적·지식적 차원 모두에서 교차 도메인 특징을 명시적으로 모델링하여 인지 부하의 영향을 포착한다. 또한 도메인 내 및 도메인 간 지식 상태 전파를 모델링하기 위한 지식 전이 모듈을 설계한다. 이러한 요인을 공동으로 모델링함으로써 LT-MKT는 학습자의 미래 수행을 더 정확히 예측할 수 있게 한다. 마지막으로, 실세계 데이터셋에 대한 광범위한 실험을 통해 제안 방법이 최신(state-of-the-art) 수준의 성능을 달성함을 보인다.

arXiv:2608.22490v2cs.CL

누가 안전성의 대가를 더 치르는가? 언어 간 안전성 정렬의 불균등한 비용 측정

Who Pays More for Safety? Measuring the Disparate Cost of Safety Alignment across Languages

안전성 정렬(safety alignment)은 모델이 인간의 가치관을 준수하도록 돕지만, 종종 응답의 유용성(utility)을 저하시킵니다. 본 연구에서는 중요하지만 충분히 연구되지 않은 질문을 던집니다. 과연 안전성 정렬은 언어 집단 간에 비용을 동등하게 부과하고 있는가? 이에 답하기 위해, 저자들은 안전성 정렬만으로 인해 발생하는 유용성 손실을 측정하는 엄밀한 프로토콜을 도입하고 이를 '안전성 비용(Safety Cost)'이라 명명합니다. 안전성 정렬된 모델과 정렬되지 않은 대응 모델 간의 직접적인 쌍별 비교를 통해, 비영어권 사용자가 영어권 사용자보다 지속적으로 더 높은 안전성 비용을 부담한다는 체계적인 불평등을 발견했습니다. 나아가 세 가지 근본적인 패턴을 규명했습니다. 첫째, 여러 언어가 이중 불이익 구간(double-penalty zone)에 속하여 더 취약한 안전성 보호와 더 큰 유용성 손실을 동시에 겪고 있습니다. 둘째, 특정 언어에서 나타나는 유용성 증가는 사실 안전성 필터가 제대로 작동하지 못한 결과입니다. 셋째, 고자원 언어(high-resource languages)조차 동일한 수준의 안전성에 도달하기 위해 영어보다 더 큰 안전성 비용을 치릅니다. 이러한 격차는 명시적인 거절뿐만 아니라 여러 차원에 걸친 암묵적인 질적 차이에서도 비롯됨을 보여줍니다. 안전성 정렬의 불균등한 영향을 정확하게 측정함으로써, 본 연구의 결과는 현재의 안전성 정렬 관행에 내재된 체계적인 격차를 드러냅니다.

arXiv:2608.22237v2cs.AI

Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents

Web-facing AI agents repeatedly access webpages, documents, repositories, and tool responses, yet reading interfaces often expose entire objects when only sparse evidence is needed. This over-reading increases token and latency costs and can dilute relevant evidence. We present SparseRead, a reusable reading control layer that governs content admission without modifying the agent model. Its Read Intent Protocol tracks evolving needs and source coverage; Adaptive Reader Backends select how evidence is located and exposed; and the Budgeted Read Gate bounds cumulative raw exposure per episode with faithful full-view fallback. A unified Reader interface supports interchangeable evidence acquisition across heterogeneous resources. Across six models, including Claude Opus 5, and five scenarios, SparseRead reduces total tokens by up to 92.9% and wall time by up to 89.0%, preserving or improving task quality in 97% settings. On Web deep-search tasks, it reduces tokens by 31.0--43.0% and task completion time by 23.0--37.0%. Ablations validate the three designs, while integration with three agent frameworks demonstrates portability.

arXiv:2608.22007v3cs.LG

트랜스포머의 커뮤니케이션 맵

The Communication Map of a Transformer

The components of a transformer communicate by writing to and reading from a shared residual stream, and the mechanistic interpretability literature has mapped these connections by hand, one circuit at a time. We present the communication map, which charts every potential communication channel from the geometry of the model's weights alone. Its unit of analysis is the whole transformer rather than the individual head or circuit. Every ordered writer-reader pair, across all 18 connection classes from head-to-head to neuron-to-neuron, receives a coupling coefficient that generalizes the composition score of Elhage et al. (2021), with an exact chance level. The census finds that 70-89% of head pairs are oriented far from chance, some coupled strongly and others actively avoiding each other. The map recovers the known induction circuits blind from the strongest head-to-head couplings and groups the heads into communities, and ablating one such community destroys the model's in-context copying. As an application, we pool the coupling coefficients of every head to identify a two-dimensional residual stream subspace, distinct from those found by activation PCA or outlier dimensions, whose deletion abolishes induction in six models up to Pythia-6.9B. We release the map, the statistical machinery, and the intervention suite.

arXiv:2608.21810v2cs.LG

MSM-Mem: 의료 AI 에이전트를 위한 범용 의료 구조화 멀티모달 메모리 프레임워크

MSM-Mem: A Universal Medical Structured Multimodal Memory Framework for Medical AI Agents

임상 의사결정은 본질적으로 경험 기반이다. 의사들은 상호작용 전반에 걸쳐 환자 병력, 멀티모달 관측, 과거 진단 경험을 종합하면서 추론을 점진적으로 정교화한다. 반면, 현재 멀티모달 대규모 언어 모델(MLLM) 기반 의료 AI 에이전트는 대체로 상태 비저장 추론 시스템으로 동작하여, 각 상호작용마다 독립적으로 결정을 생성하며 경험적 지식을 유지하거나 내재화하지 않는다. 이러한 불일치는 사용을 통한 추론 신뢰도의 점진적 향상과 실제 임상 워크플로에서의 종단적 환자 맥락 적응 능력을 제한한다. 본 연구에서는 누적된 임상 경험을 통해 의료 AI 에이전트가 진화할 수 있게 하는 에이전틱 메모리 프레임워크인 Medical Structured Multimodal Memory(MSM-Mem)를 제안한다. MSM-Mem은 이질적인 임상 경험을 의미, 에피소드, 시각 메모리로 구조화하고 추론 중에 이를 점진적으로 갱신하여, 에이전트가 과거 경험을 검색해 현재 추론에 반영하고 시간에 따라 의사결정을 점진적으로 정교화할 수 있게 한다. MoE-LLaVA 백본에 대한 평가는 지속적 사용을 통해 추가 이득이 관측되는 일관된 성능 향상을 보여준다. 전반적으로 MSM-Mem은 임상의가 시간에 걸쳐 실무에서 배우는 방식과 유사하게 추론 역량을 진화시킬 수 있는 의료 AI 에이전트를 향한 실행 가능한 경로를 제공한다.

arXiv:2608.20271v2cs.AI

Catching the Rug: Early Prediction of Fraudulent Memecoins on Solana via Machine Learning

The rapid proliferation of memecoins on blockchain platforms has increased the risk of fraudulent activities, particularly rug pulls. While previous studies have focused on Ethereum-based tokens, this paper shifts the spotlight to Solana, the leading blockchain for memecoins by trading volume and token count. Unlike Ethereum, where rug pulls often exploit smart contract backdoors, Solana memecoin rug pulls are predominantly driven by liquidity manipulation and social dynamics. This research pioneers large-scale rug pull early detection in the Solana ecosystem by assembling a dataset of 6.4 million tokens over 7 months. Market analysis reveals that a vast majority of these memecoins exhibit rug pull characteristics within one hour of launch, highlighting the urgency of short-horizon prediction. Despite the absence of code-level features, we demonstrate that classic machine learning models, particularly Gradient Boosting (XGBoost), achieve robust performance in detecting potential rug pulls using only the first 5 minutes of trading data. Furthermore, we evaluate cross-platform generalization between PumpFun and Raydium, revealing that multi-source data fusion significantly mitigates domain shift and improves detection reliability. This study advances the understanding of DeFi fraud on high-throughput chains and provides a practical framework for protecting investors.

arXiv:2608.19861v2cs.AI

PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents

Customer-service LLM agents must follow organizational policy when acting on a user's behalf. Compliance failures arise from either forbidden actions, such as granting an ineligible change, or omitted procedural requirements, such as identification or confirmation. Runtime safeguards can intervene on risky actions, but action-local checks do not guide an agent through a multi-step procedure. Workflow-following systems support prescribed process execution, but primarily target workflow completion rather than safeguarding agent behavior. PolicyGuide instead compiles each domain policy into a workflow graph and invokes a proactive verifier at user-turn boundaries. From persisted graph state, the verifier reconciles open requests and returns step-specific remediation along a policy-compliant path. Across the $\tau^2$-bench airline, retail, and telecom domains with a GPT 5.4 agent and verifier, PolicyGuide raises mean $\mathrm{Pass}^4$ from $0.42$ to $0.62$, with the largest gain on telecom ($0.19$ to $0.61$), the most workflow-structured domain. The same workflows transfer to Claude Sonnet 4.6 and Gemini 2.5 Pro agents. Complementary evaluations find the lowest observed attack-success rate under adversarial users and the strongest procedural compliance in an independently validated workflow-level audit.

arXiv:2608.18495v2cs.LG

Physics-Unrolled Neural Operator for Accelerating High-Fidelity Wireless Ray Tracing

Site-specific wireless design relies on ray tracing to predict how signal strength varies across a floorplan, but a converged indoor radio map takes tens of seconds to simulate and planning a single building requires thousands of them. A learned surrogate can remove most of that cost, yet training one runs into a second problem: high-fidelity labels are exactly what is too expensive to produce at scale, so the affordable labels come from finite-ray simulations that carry residual Monte Carlo noise. We address this with RayNO, a physics-unrolled neural operator that reconstructs the converged output of a ray tracer from one cheap low-fidelity run and scene priors. Rather than treating a radio map as a generic image, RayNO refines the field through three stages matched to reflection, diffraction, and scattering, each supervised by a label of matching physical complexity. We prove that when the label error is conditionally unbiased the learned operator can end up closer to the high-fidelity reference than the labels it was trained on, and we verify that condition directly in the simulator using repeated independent ray seeds. Across diverse floorplans, RayNO outperforms image-to-image baselines, wireless learning models, monolithic neural operators, and classical multi-fidelity regression on both image-quality and wireless deployment metrics, while reducing the cost of a high-fidelity map from roughly 28 s to 0.11 s.

arXiv:2608.17583v2cs.CL

Auditing Exposure to Harmful Content on TikTok using Multimodal Language Models: A Cross-National, Age-Stratified Study

Online video platforms can expose young users to harmful content, but independent audits remain difficult because video annotation is costly and moderation judgments vary across languages. We audit TikTok in France, Italy, and Sweden with sockpuppet accounts representing four age personas (13, 16, 19, 40), collecting 36,971 videos from passive For-You-page scrolling and active sessions that scroll, search for harm keywords, and scroll again. To scale annotation, we validate four multimodal LLMs against native-speaker labels on a 300-video reference set. Gemini 2.5 Flash with eight sampled frames plus text performs best (aggregate kappa = 0.42), at half the per-call cost of native-video upload, and we apply it to a 10% sample for approximately \$50 in total API spend across both modalities. Keyword search returns 35-56% harmful content, a 1.5-7.5x increase over the scrolling baseline in ten of twelve country-age combinations; the spike is temporary and flattens the age differences observed in France and Sweden. Under passive scrolling, Italy has the highest harm rate at every age, with Italian age-19 reaching 48.6%. Overall, MLLM-based auditing offers a scalable approach for cross-national youth-safety audits, while provider safety filters (1.1% refusal rate) under-count the most explicit harms.

arXiv:2608.17253v3cs.LG

Co-RL: 다중 에이전트 RL에서 다양한 코호트로부터 무감독 추론의 창발

Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL

Reinforcement Learning (RL) has emerged as a powerful approach for improving reasoning in language and vision-language models, yet its strongest successes still depend heavily on ground-truth supervision (e.g., verifiable reward). Such annotations are costly to obtain and become increasingly scarce as reasoning capabilities advance beyond what humans can reliably evaluate. Self-rewarding RL reduces this dependence by enabling models to derive reward signals from their own completions. However, training solely on self-generated feedback can reinforce existing biases, reduce response diversity, and ultimately lead to homogenized responses and training collapse. In this work, we show that unsupervised reasoning can emerge through cooperative multi-agent training. We introduce Co-RL, a framework in which multiple decoupled models, sharing no parameters, are simultaneously optimized through RL using rewards derived from the other agent. We further show that increasing cohort diversity, through heterogeneous model families, sizes, and rephrased training samples, reduces the correlated errors that drive self-reinforcing feedback loops. This diversity consistently improves reasoning performance and mitigates training collapse. Without any ground-truth labels, Co-RL improves the base models by 3.0-7.9% on average across seven benchmarks and by 1.6-4.4% across four multimodal benchmarks. It also outperforms the prior state-of-the-art multi-agent RL method CoMAS by 4.0% and the self-rewarding vision-language RL method MM-UPT by 2.0%. Our code is available at https://github.com/Diverse-CoRL/Co-RL and website at https://williamium3000.github.io/Co-RL.

arXiv:2608.17162v2cs.LG

OraclePhys: A Systematic Framework for LLM Fine-Tuning on Structural Mechanics

What fine-tuning teaches a model is usually diagnosed after the fact. We make it an experimental variable. OraclePhys has three parts. OraclePhys-Bench is a structural-mechanics benchmark whose finite-element oracle grades every answer and counterfactual edit. OraclePhys-30K holds seven answer forms over byte-identical structure descriptions. A controlled training study compares the forms and verifier roles. Two findings follow. First, within this task family the answer form of the label determines what fine-tuning teaches. Ranking supervision lifts held-out localization of the most-deforming story on taller structures from 0.25 to 0.86, a peak-value label yields at best a partial predictor, and a pass/fail label none. This ranking-versus-peak gap replicates on a second domain, a second model family and a paraphrased surface. Second, the gap between our standard SFT and GRPO runs is mostly a recipe effect. With learning rate and step budget matched, the two reach held-out accuracies within 0.03 of each other. At higher rates, GRPO falls below its start while SFT does not. Both also teach the model to apply edits given as instructions. The trained 8B matches a specialist and exceeds a frontier LLM. Both domains are linear, steady-state systems with synthetic text inputs.

arXiv:2608.17070v3cs.LG

Certified but Private: Scalable Zero-Knowledge Proofs for Neural Network Guarantees

With the growing deployment of machine learning models, formal guarantees of the robustness and fairness of these models have become increasingly important in safety-critical and legal-compliance settings. However, model parameters are often commercial secrets that cannot be disclosed to auditors or end users. To this end, we present PANDA, a scalable system that uses zero-knowledge proofs (ZKPs) to prove the robustness and fairness properties of a model without revealing its private parameters. PANDA is built on top of CROWN, an efficient robustness certification framework that is used in many state-of-the-art formal verification tools for neural networks. The core contribution of PANDA is a novel algorithm for proving linear relaxation bounds for non-linear activation layers, yielding simple, lightweight proofs. Remarkably, our implementation can generate proofs of local robustness for neural networks with more than 2.9M parameters in 5 minutes, and can verify them in 10 seconds. Prior ZKP-based robustness system rely on exponential-time algorithms that cannot scale to nontrivial networks. In contrast, PANDA scales polynomially in the number of neurons in a network, allowing us to support neural networks 4 orders of magnitude larger than previous approaches with significantly reduced prover overhead.

arXiv:2608.16347v2cs.CL

대규모 언어 모델 간 활성화 상태의 아키텍처 종속적 인과 전달

Architecture-Dependent Causal Transfer of Activation States Across Large Language Models

AI 시스템 간 직접 통신은 중간 계층으로 자연어에 의존하며, 인코딩/디코딩 오버헤드, 토큰 비용, 지연을 초래한다. 우리는 학습된 투영(projection)을 통해 서로 다른 대규모 언어 모델(LLM) 아키텍처 간에 내부 활성화 상태가 대신 인과적으로 전달될 수 있는지를 세 가지 수준에서 평가한다: 표현 유사도, 투영된 상태로부터의 교차 모델 검색, 생성 중 활성화 주입을 통한 종단(end-to-end) 인과 전달. 아키텍처적으로 다양한 네 가지 오픈 가중치 모델(Qwen2-0.5B, Phi-3-mini, Mistral-7B, FLAN-T5-base)을 사용하여, 학습된 모델에서의 표현 정렬(alignment)이 무작위 초기화 null 기준선을 초과하며, 순위 기반 지표(mutual k-nearest-neighbour alignment)로 가장 잘 포착되고, centered kernel alignment(CKA)나 Procrustes 분석보다 활성화 크기 이상치(outlier)에 더 강건함을 발견한다. 학습된 투영 네트워크는 세 가지 인과적 디코더 전용 모델 쌍에 대해 보류 평가셋에서 정확한 타깃 모델 표현을 우연 수준보다 훨씬 높게 검색한다(top-1 정확도 45–50% 대 우연 5%)하지만, 인코더 기반 FLAN-T5에서는 우연 수준에 머문다. 생성 중 투영된 활성화를 타깃 모델에 주입하면 세 디코더 전용 쌍 중 하나에 대해서만 검색 기반 출력 유사도에 통계적으로 유의하고 사전 등록(pre-registered)된 인과 효과를 낸다(Qwen2-0.5B에서 Phi-3-mini로: 23.3% 대 부정 대조 하 0.0%, p=0.047, FDR 보정); hidden-state 수준에서 유사한 표현 정렬에도 불구하고 Mistral-7B를 타깃으로 하는 두 쌍에서는 그러한 효과가 나타나지 않는다. 우리는 이러한 결과를 의미(meaning)가 아닌 표현 매개체(representational vehicle)의 인과적 전달에 대한 증거로 해석하며, 현재 구현된 대로 LLM 간 종단 활성화 상태 전달은 보편적이기보다 아키텍처에 의존한다고 결론짓는다.