LLM-based coding agents have advanced rapidly on single-process SWE tasks, with frontier models now clustering in the high-70s on SWE-bench Verified. Distributed-system debugging, however, remains an under-explored regime: bugs span processes, nodes, and protocol interactions, with root causes rarely recoverable from source alone and brute-force exploration intractable across non-deterministic interleavings. This leaves two gaps in LLM and agent evaluation: no code-repair benchmark targets distributed-system bugs, and no controlled study isolates how much externally provided debugging context changes agent success on them. We introduce DDBench, a code-repair benchmark of 60 historical bugs mined from 13 open-source distributed systems, partitioned into three reasoning-scope tiers. DDBench evaluates every case under two matched conditions: a symptom-only condition where the agent receives only the bug symptom and repository, and a context-augmented condition where it additionally receives bounded debugging context (logs, traces, runtime state, and targeted code-investigation notes), isolating the effect of debugging context from model capability. The evaluation of 10 LLMs on DDBench reveals several findings. First, distributed debugging exercises a reasoning dimension that single-process benchmarks do not surface: models' pass rates span 45 pp, and pairwise bootstrap separates 6 of 15 top-tier model pairs at p < 0.05 on DDBench's hardest case-set. Second, bounded debugging context lifts aggregate pass rate by +19.0 pp, and the lift is asymmetric: weaker models gain more in pass rate, while stronger models gain more in efficiency. Third, debugging context requires careful curation, as even faithful debugging context can anchor a strong model on the wrong layer of the system when its observation sits far from the fault.
Residual connections are a fundamental component of transformer architectures, yet the roles of the attention and feed-forward residual pathways remain poorly understood when considered independently. This paper presents a reproducibility study of partial residual ablations in Pre-LN GPT-style transformers trained at two scales (10M and 124M parameters). I compare four architectural configurations by selectively removing the attention residual connection, the feed-forward residual connection, or both. At 10M scale, a controlled 8-seed deterministic sweep shows a clear asymmetry: removing the attention residual (FFNOnly) reaches the No-Residual collapse regime (3.350 +/- 0.002), whereas removing the feed-forward residual (AttnOnly) remains well separated from it (1.580 +/- 0.003). A subsequent controlled 124M sweep establishes that the direction of this asymmetry persists under matched seeds: across five seeds each, AttnOnly reaches 6.037 +/- 0.355 versus 7.569 +/- 0.001 for FFNOnly, with no overlap between the observed ranges. AttnOnly nevertheless remains substantially degraded relative to FullResidual (4.577 +/- 0.012) and exhibits markedly greater seed sensitivity at 124M. During the investigation, I identified and corrected an experimental measurement confound in runtime gain scaling and retained an intermediate reproduction failure rather than excluding it. I propose cross-position routing through self-attention as a falsifiable hypothesis for the observed asymmetry; no clean causal test of this mechanism has yet been completed. To support reproducibility, I release the source code, experiment configurations, training logs, and experimental results, including intermediate non-reproducing runs.
Modern stochastic predictors can model rich, multi-modal outcome distributions. However, this expressive power comes with challenges in ensuring robust predictions $-$ a critical requirement in safety-critical domains. Randomized smoothing is a leading technique for improving robustness, particularly against adversarial perturbations. Yet, in stochastic multi-modal regression settings, randomized smoothing often fails due to mode collapse, yielding averaged predictions that do not reflect the underlying distribution. To address this limitation, we propose clustered $\alpha$-smoothing, a framework that (1) partitions noisy samples using an arbitrary clustering algorithm, (2) applies $\alpha$-smoothing locally within each cluster, and (3) combines the resulting predictions into a mixture distribution. By interpreting the smoothing distribution as a mixture of $\alpha$-smoothers, we derive a lower bound on the probability that the smoothed prediction lies within a union of compact regions corresponding to distinct modes. We empirically evaluate our framework on two benchmarks, demonstrating substantial improvements over state-of-the-art methods. In stochastic trajectory prediction on a driving simulator dataset, our approach achieves, on average, a $27\%$ lower Wasserstein distance to the ground-truth distribution compared to $\alpha$-smoothing. In quadrotor control, where modes correspond to distinct feasible paths to a target, our method reduces the collision rate by $90\%$ relative to the state-of-the-art randomized smoothing.
Air traffic controllers perceive traffic complexity through the radar display, suggesting that a computer vision model operating on the same imagery may provide a natural architecture for modeling controller-perceived complexity; however, whether radar imagery is a viable input format for deep learning vision models remains unclear. Unlike natural images, radar images are extremely sparse and self-similar, consisting primarily of a black background and a few visually identical aircraft blobs, while small changes in aircraft positions can substantially alter sector-level complexity. To test whether a vision model can capture these operationally important differences, we encode each traffic situation as a position image supplemented by five channels representing aircraft state variables, including heading, speed, and altitude, and train a Vision Transformer (ViT) to regress four intrinsic complexity components derived from pairwise geometric relations among aircraft. The model achieves $R^2 > 0.96$ for all four components on test data, and a one-aircraft removal perturbation study shows that the change in the model's estimate is proportional to how much the removed aircraft contributed to sector complexity, rather than the same for every removal. These results demonstrate that, despite its atypical visual characteristics, the proposed image-based representation is a viable input format for air traffic complexity modeling.
Low-resource automatic speech recognition (ASR) often relies on cross-lingual transfer from higher-resource languages due to limited availability of transcribed speech. However, selecting an effective donor language is a nontrivial issue, particularly in settings with spontaneous speech, linguistic variation, and uneven resource availability. We present DonorRank, a learning-to-rank framework that uses linguistic and dataset-level features along with previously observed transfer outcomes to predict promising donor languages for zero-shot ASR. We evaluate DonorRank on two multilingual speech corpora of Indic and African languages. DonorRank produces informative donor rankings across both collections, outperforming common heuristics in the more closely related Indic setting while remaining competitive in the more linguistically diverse African setting. Our analysis reveals that effective donors are not always the closest linguistic relatives and that the most informative predictors vary across language collections. These findings highlight the value of learned donor selection for prioritizing transfer experiments in low-resource ASR. We will also release all of our evaluation results, feature matrices for all languages, scripts for training the DonorRank framework, and pairwise transfer performance matrices for our datasets.
Conway's 99-graph problem asks whether a strongly regular graph with parameters srg(99,14,1,2) exists. We develop two complementary lines of attack. Fixing one vertex, the conditions lambda=1 and mu=2 force its neighbourhood to be a perfect matching and determine every edge between that neighbourhood and the remaining vertices. For (99,14,1,2), the unresolved part is therefore a constrained 12-regular graph on 84 vertices. We encode this reduction in CP-SAT and validate it by recovering the unique srg(9,4,1,2). We also prove by exhaustive enumeration that no circulant graph on Z/99 satisfies more than 68.0% of the CAISc constraints, and we give a validated orbit formulation for prescribed automorphisms. We then study the partial-score search problem. Fourteen human-designed search configurations reached at most 69.43%. Separately, we supplied the scoring function to an evolutionary program-search system. It produced a degree-preserving 4-vertex-switch tabu search whose best verified artifact scores 70.73%. The generated move differs from those used in our own searches and crosses a plateau that was stable under them. A tabu search restricted to graphs with a prescribed symmetry group, first Z/33 and then its subgroup of order 11, reaches a 14-regular graph scoring 76.67%. These results do not resolve the existence problem, but they reduce the exact search space and improve the best verified partial construction found in our experiments.
Decompose-then-verify pipelines (FActScore-style fact-checkers, hallucination detectors) first split a passage into atomic claims and treat this step as neutral preprocessing. We ask whether a decomposer can substitute its own parametric belief for what the passage says, producing a claim that contradicts its source, and call this Decomposition-Induced Context-Memory Conflict (DI-CC). On biographies with a deliberately altered year, Qwen2.5-7B-Instruct does so in 4.3% of claims when the instruction licenses self-correction, but in only 0.2-0.5% under a neutral instruction, a FActScore-style few-shot prompt, a "use only the passage" instruction, a formatting instruction, or a knowledge-filling instruction (an upper bound: in the neutral and FActScore-style runs only 1 of 13 flagged claims is a real override); context-aware decoding removes only half. DI-CC here is therefore instruction-induced. We then test whether a linear probe trained only on classical context-memory conflict data (NQ-Swap) detects DI-CC zero-shot. Its AUC is high (up to 0.94 against legit claims, 0.98 under a position- and instruction-matched control), but random directions reach AUCs between about 0.1 and 0.9 on these contrasts, probes fit directly on the claims reach 0.96-1.00 everywhere (even on a control that separates only the text variant), and after multiplicity correction the NQ-Swap probe beats random directions only at the last layer, and only for two of three random-direction families. A direct NLI check between source and claim reaches AUC 0.998. High probe AUC on such small contrasts is not evidence of a shared mechanism without nulls, and two pitfalls, pair leakage in probe training and duplicated BOS tokens, can change such conclusions.
Detecting a traffic anomaly does not establish whether a video-language model can explain what happened, localize it in time, or identify its causes. We introduce TAR (Traffic Anomaly Reasoning) and TAR-Bench, paired resources for training and evaluating these complementary capabilities across 10 tasks spanning question answering, temporal reasoning, and scene understanding. TAR contains 44,040 automatically generated annotations with chain-of-thought traces for 3,670 CCTV videos from eight public datasets. TAR-Bench provides 960 human-curated annotations for 80 held-out clips from 17 public YouTube videos. Evaluation of eleven vision-language models and analysis of fine-tuned outputs reveal temporal, causal, and descriptive errors not captured by closed-form question-answering accuracy. Training on all 10 tasks raises mean benchmark scores from 34.3 to 55.7 for Cosmos-Reason2-8B and from 30.9 to 53.9 for Qwen3-VL-8B-Instruct. These results support joint training across complementary tasks, while highlighting persistent limitations in temporal precision and causal attribution. TAR and TAR-Bench serve as the official training and in-domain evaluation resources for AI City Challenge 2026 Track~3. The dataset is available under CC-BY-4.0 at https://huggingface.co/datasets/nvidia/PhysicalAI-Traffic-Anomaly-Reasoning.
All-in-one image restoration is a unified low-level vision task that aims to effectively recover high-quality images from inputs degraded by various types and levels of corruption using a single model. Recent works have achieved remarkable progress by learning degradation-adaptive prompts or network architectures. However, these methods typically apply a uniform restoration strategy across the entire image, neglecting the fact that different regions may suffer from distinct degradation types and varying degrees of severity. In contrast, we propose to perform restoration at the pixel level, thereby enabling more fine-grained and precise control over the restoration process. Specifically, we present MGN-AIR, a novel pixel-level restoration framework for all-in-one image restoration. Our approach first learns to estimate a pixel-level visual prompt. Then, it leverages both textual and visual prompts to provide global and local degradation cues, guiding the model on where to look and how to restore at each pixel. We conduct extensive experiments on multiple all-in-one image restoration benchmarks, covering a wide range of tasks including denoising, deraining, deblurring, dehazing, desnowing, and low-light enhancement. Experimental results demonstrate that our proposed method consistently and significantly outperforms existing approaches.
Monocular 3D object detection spans two regimes: closed-set detectors operating within a fixed category vocabulary, and open-vocabulary detectors that localize arbitrary categories by leveraging depth foundation models for 3D geometry. We find that current depth foundation models, despite their strong zero-shot generalization, lack the object-level precision 3D detection demands: substituting a state-of-the-art depth foundation model for a strong detector's predicted depth degrades accuracy. Rather than pushing detectors or depth models to be more accurate end-to-end, we treat object-level depth refinement as a stand-alone task and present RefineAny3D, a vision-language model that corrects depth without ever predicting a numerical value. Our key insight is that depth error has a direct visual signature in image space: when projected onto the image, a correctly placed box tightly encloses the object, while a too-far box projects too small and a too-close box projects too large. Depth refinement thus reduces to a visual alignment problem rather than a metric regression problem, which we instantiate by extending the VLM's vocabulary with action tokens that replace numerical depth output with categorical decisions, and by supervising the model on a large-scale chain-of-thought dataset that grounds each decision in explicit visual evidence. Applied as a single post-hoc step, RefineAny3D delivers consistent gains across closed-set detectors, open-vocabulary detectors, and 3D auto-labeling tools, and generalizes to novel categories, scenes, and cameras without retraining. Code is available at https://github.com/alanzhangcs/RefineAny3D.
대규모 생성 모델을 활용한 이미지 초해상도(SR)는 최근 뛰어난 인지적 품질을 달성했으나, 저해상도(LR) 관측에 대한 충실도를 유지하는 것은 여전히 어려운 과제로 남아 있습니다. 특히, 잠재 표현(latent representation)을 기반으로 구축된 확산 트랜스포머(DiT)는 VAE의 압축 병목 현상으로 인해 세밀한 공간 정보가 약화되어 입력 이미지에 취약하게 그라운딩(grounding)된 세부 표현이 환각(hallucination)되는 치명적인 한계가 있음을 관찰했습니다. 본 연구에서는 표현 관점에서 생성형 SR을 재검토하고, VAE 압축 이전의 LR 관측 픽셀 증거를 보존하여 복원 전반에 걸쳐 재사용하는 픽셀 그라운딩 초해상도(PGSR) 프레임워크를 제안합니다. PGSR은 압축된 잠재 조건에만 의존하는 대신, 업샘플링된 LR 이미지에서 VAE 이전(pre-VAE) 픽셀 증거를 추출하여 두 단계에서 재사용합니다. 첫째, 조건 측 궤적 유도(Condition-Side Trajectory Guidance)는 LR에서 파생된 픽셀 증거를 잠재 LR 조건과 융합하여 잠재 복원 궤적을 안내합니다. 둘째, 디코더 측 픽셀 그라운딩(Decoder-Side Pixel Grounding)은 동결된 VAE 디코더에 다중 스케일 픽셀 특징을 주입하여 LR 관측 단서로 최종 렌더링을 그라운딩합니다. 사전학습된 대규모 DiT 모델을 효율적으로 적응시키기 위해, 잠재 오토인코더와 주 플로우 매칭(flow-matching) 백본은 동결 상태로 유지하고 경량 복원 모듈만 학습합니다. 나아가 고해상도 효율성과 확장성을 개선하기 위해 효율적인 로컬 윈도우 어텐션 변형을 연구합니다. 광범위한 실험을 통해 PGSR이 사실성과 충실도 간의 절충(realism--fidelity trade-off)을 개선하며, 기존의 잠재 생성형 SR 방식보다 더 충실하고 시각적으로 설득력 있는 결과를 생성함을 입증합니다.
Large Language Model (LLM) agents are increasingly deployed in multi-agent social settings where they must cooperate, negotiate, and adapt to other agents. Measuring these social skills is hard because most social interactions offer no objective ground truth. Evaluations often fall back on LLM judges, which are costly, subjective, and noisy. We introduce SocialGym, an environment of 22 multi-agent social games, from matrix games to hidden-role deduction, with verifiable outcomes determined by game rules. We benchmark ten LLMs across all games and roles using an Elo tournament. Our results show that no model is best in every game, and rankings shift across game families and sides. Secret keeping correlates more strongly with theory-of-mind and privacy benchmarks than with general knowledge benchmarks. Half of the models give away their side's secret in Chameleon and Spyfall, and the smallest models are not reliably better than chance at voting for hidden opponents in most hidden-role games. Finally, we examine whether models can improve their play by reflecting on self-play and writing strategy playbooks, finding mixed results across models and games. SocialGym provides verifiable measures for diagnosing LLM social capabilities across settings and roles.
A single model scale challenges the flexibility of a production retrieval system: some settings need it faster, others need a smaller index, and the right trade-off changes with the workload. In the context of information retrieval (IR), a transformer-based model can be made smaller in three ways --- using fewer layers, passing fewer tokens through the upper layers, or producing a shorter embedding --- and each way saves a different resource. These options have been studied in isolation, each as its own method with its own code and training setup, which makes them hard to combine or adapt to a new model. In this work, we present Tevatron-Elastic to bring all three under one unified abstraction: a single object names any size the model can run at, and a short schedule lists the sizes to train. Training then produces one checkpoint that serves all of those sizes, and at deployment the user has the liberty to pick any of them. The same abstraction covers both retrievers and rerankers and both encoder and decoder models, as it works through interfaces that Hugging Face transformers already expose; adding a new backbone can be a configuration change rather than a model rewrite. Prior methods --- Matryoshka embeddings, early exit, 2D Matryoshka (Starbucks), and layerwise token compression (LTC) --- become special cases of our unified abstraction. With Tevatron-Elastic, we also propose a new elastic algorithm Matryoshka~LTC (MLTC) to jointly train several token-compression ratios in one retriever checkpoint. To validate our framework, we train 20 checkpoints across three backbones and two tasks: the quality curves are smooth, full-point quality remains close to that of models trained for a single size, and a controlled study confirms the wall-clock speedups. We release the framework and all checkpoints to support elastic retrieval research.
해석 가능성(interpretability)은 종종 능력(capability)에 대한 대가로 취급된다. 언어 모델은 불투명한 시스템으로 학습된 뒤 사후에 설명되며, 그 방법들의 신뢰성을 확립하기 어렵다. 본 연구에서는 이 전제에 도전한다. 모델을 역공학하는 대신, 해석 가능성을 학습 파이프라인의 제약으로 두고 언어 모델링 목적과 함께 최적화한다. 컴퓨트 규모가 세 자릿수(order of magnitude)에 걸쳐 변하는 범위에서, 자기회귀(autoregressive) 및 확산(diffusion) 언어 모델 모두에 대해 해석 가능성은 능력과 상충하기보다 함께 스케일링된다. 놀랍게도, 규모가 커질수록 모델 표현은 더 분리(disentangled)되고 사람이 이해 가능한 개념과 더 정렬(aligned)된다. 우리는 Steerling-8B—인과적 어텐션 마스크를 사용하는 확산 언어 모델—로 학습 시점 레시피를 구현한다. 생성된 토큰 그룹마다 Steerling-8B는 출력을 관련 입력 토큰, 사람이 이해 가능한 개념, 학습 데이터에 귀속(attribution)한다. 이를 통해 폐루프(closed-loop) 개입이 가능해진다. 개념 또는 특징 귀속으로 출력을 진단하고, 유사한 학습 데이터를 검색하며, 재학습 없이 개념 스티어링(concept steering)으로 행동을 교정할 수 있다. Steerling-8B는 상당히 2–16배 더 많은 컴퓨트로 학습된 공개 동급 모델들과 경쟁력을 유지하며, 다른 스케일링 패러다임을 시사한다. 해석 가능성은 학습에 설계로 내재될 수 있고, 규모와 함께 개선될 수 있다.
Large language models (LLMs) are increasingly involved in scientific discovery, yet it remains unclear whether they can support complex real laboratory science. Here we introduce Science Edge Evaluation (SEE), a multimodal benchmark of expert-curated questions grounded in peer-reviewed literature and experimental practice in chemistry, biology, and materials science. Evaluation of 28 multimodal large language models (MLLMs) shows that even the best-performing model reaches only 54.9% accuracy. Moreover, general-purpose models outperform science-specialized models on average. In the visual-agent evaluation, the use of tools increases the best accuracy to 56.1%. Tool use can expand the information available to models, but more information does not necessarily lead to reliable scientific reasoning. The key challenge is whether models can manage tool-derived information within the boundaries of the original experimental evidence. Together, these findings reveal that current MLLMs still cannot reliably make justified and evidence-bounded inferences from experimental results, which is an essential capability in real scientific discovery. Bridging this gap requires MLLMs to transition from explaining established scientific concepts to deriving novel and evidence-based insights from experimental data.
과매개변수화된 분류(classification) 문제에서는 기저 분포가 선형적으로 분리되지 않더라도 학습(training) 데이터는 선형 분리가 가능할 수 있습니다. 이러한 설정에서 로지스틱 손실(logistic loss)에 대한 경사하강법(GD, gradient descent)은 방향 측면에서는 최대 마진(max-margin) 보간 분류기로 수렴하지만 노름(norm)은 발산하며, 이때의 암묵적 바이어스(implicit bias)는 통계적으로 준최적(suboptimal)일 수 있습니다. 본 연구에서는 조기 종료(early stopping)가 이러한 준최적성을 극복할 수 있음을 보여줍니다. 즉, 레이블 반전 노이즈가 있는 가우시안 혼합 모델(Gaussian mixture model)에서 적절한 오라클 시점에 중단된 경사하강법은 다항 및 지수 스펙트럼 감쇠를 포함하여 빠르고 연속적인 감쇠를 갖는 공분산 스펙트럼에 대해 미니맥스 최적의 초과 0-1 위험(excess zero-one risk)을 달성합니다. 본 분석은 조기 종료된 반복해에 대한 정밀한 상한과 임의의 분류기에 대한 일치하는 통계적 하한을 결합하여 최적 수렴률을 도출하며, 이는 실험을 통해 검증됩니다. 핵심적인 기술적 기여는 초과 로지스틱 위험을 초과 0-1 위험으로 변환하는 새로운 캘리브레이션 결과입니다. 이는 레이블 반전 노이즈로 인해 발생하는 모델 오지정(model misspecification)을 처리하고 표준 바운드에 나타나는 제곱근 비율을 제거합니다. 또한 선형 보간기에 대한 하한을 규명하여, 보간 방식이 동일한 초과 위험을 달성하기 위해 조기 종료에 비해 지수적으로 더 많은 샘플을 필요로 할 수 있음을 보입니다.
대규모 언어 모델(LLM)이 소프트웨어 시스템 및 AI 서비스에 통합됨에 따라, 효율적인 LLM 서빙이 소프트웨어 공학의 중요한 관심사로 떠오르고 있습니다. 추론 과정에서 지연 시간과 처리량을 유지하면서 계산, 메모리, GPU 자원 및 실행 최적화가 요구되기 때문에 LLM 서빙은 매우 까다롭습니다. 기존 연구에서 다양한 LLM 추론, 최적화 및 서빙 기법과 프레임워크를 제안해 왔으나, 이러한 요소들이 실제 현업에서 어떻게 도입되고 있는지에 대해서는 알려진 바가 거의 없습니다. 본 연구에서는 오픈소스 소프트웨어 시스템에서 LLM 서빙 프레임워크 및 서빙 기법의 활용 현황을 조사합니다. 구체적으로 vLLM, SGLang, TensorRT-LLM, LMDeploy, FlashInfer라는 다섯 가지 LLM 특화 프레임워크를 식별하고 분석합니다. 이러한 프레임워크와 기법이 단독으로 또는 조합되어 어떻게 채택되는지, LLM 범주에 따라 채택 양상이 어떻게 다른지, 그리고 저장소별로 의도, 초점, 사용 사례 및 아키텍처 설계가 어떻게 달라지는지 살펴봅니다. 연구 결과에 따르면 vLLM이 대중성과 도입 측면에서 가장 두드러진 프레임워크이며, 병렬 계산, 메모리 관리, 네트워크 가지치기(pruning)가 가장 빈번하게 사용되는 서빙 기법 범주로 나타났습니다. 다중 프레임워크 사용은 제한적이어서 개발자들은 주로 단일 서빙 프레임워크에 의존하는 경향을 보이지만, 프레임워크를 결합하여 사용하는 경우에는 서빙 스택 전반에 걸쳐 상호 보완적인 기능을 연계하고 있었습니다. 프레임워크 도입 양상은 모델 계열, 모달리티, 모델 크기, 도메인 특화 여부 및 배포 환경에 따라 다양하게 나타났습니다. 저장소 수준의 분석에서는 LLM 서빙 프레임워크가 강화학습(RL) 기반 추론, 멀티모달 생성 및 이해, 마이크로서비스, 클라우드 인프라 등을 포함한 다양한 애플리케이션과 아키텍처를 지원하고 있음을 보여줍니다. 종합적으로 본 연구는 실제 환경에서 LLM 서빙 프레임워크의 채택 현황을 대규모로 실증 규명하여, LLM 시스템을 다루는 연구자, 프레임워크 관리자 및 실무자들에게 유용한 통찰을 제공합니다.
Time series forecasting is fundamental to temporal data mining, yet numerical models often struggle to account for changing contextual factors. A key challenge is determining when contextual evidence should guide forecast revisions while maintaining consistency with temporal patterns and domain knowledge. To address this, we propose CastFSR, a fast--slow--reflect agentic framework for context-aware time series forecasting. Fast-thinking forecasting analyzes numerical patterns, selects a suitable forecaster, and constructs a data-driven forecast prior. Slow deliberative reasoning incorporates relevant contextual evidence to selectively refine this prior, while reflective evaluation checks the revised forecast against temporal patterns, contextual evidence, and domain constraints. A shared state maintains forecasting priors, contextual evidence, intermediate forecasts, and feedback across stages. CastFSR supports both training-free LLM coordination and two-stage post-training for compact LLM coordinators. Workflow-aware trajectory distillation (WATD) learns state-conditioned tool use and forecasting decisions from teacher trajectories, while trajectory-level policy optimization (TLPO) further improves the coordinator through reinforcement learning guided by forecasting accuracy and workflow feedback. Experiments on multiple real-world datasets under an offline evaluation protocol using historical context demonstrate competitive performance against representative baselines. Our code is available at https://github.com/Xiaoyu-Tao/CastFSR.
The widespread adoption of proprietary Large Language Models (LLMs) via closed APIs poses an interpretability challenge. While white-box and gray-box (logit-dependent) attribution methods require internal model access, post-hoc attribution under a strictly black-box, text-only generative regime remains largely unexplored. To address this, we propose a model-agnostic, post-hoc attribution interpreter operating at the sentence level. We first train a transformer-based Energy-Based Model (EBM) as a surrogate to score the consistency between prompt and response sentences based on observed LLM generations. To prevent energy collapse and avoid a focus on superficial artifacts, we design contrastive negative samplers that evaluate partial counterfactual dependencies across prompt and response segments. This learned energy landscape then guides the training of a lightweight interpreter network. After model-specific training, the interpreter operates as a standalone tool; it estimates the influence of prompt sentences on a user-specified output sentence without requiring runtime API queries to the target LLM. Comprehensive experiments demonstrate that our EBM surrogate captures nuanced generative dependencies beyond surface-level heuristics, while the interpreter achieves competitive attribution quality with parameter-efficient, $\mathcal{O}(1)$ inference.
Tool-using large language model (LLM) agents produce long, multi-turn trajectories, making gradient-based post-training memory-intensive. Evolution strategies (ES) enable memory-efficient full-parameter post-training without backpropagation and can eventually match the performance of gradient-based reinforcement learning (RL). However, resource-constrained settings typically offer only a few GPUs, so the high GPU-hour requirements of ES translate into prohibitively long training times. To address this, we introduce Cooperative Parameter-subspace Evolution Strategy (CoPES), a cooperative coevolutionary method that decomposes the full parameter space into lower-dimensional subspaces and searches over them cooperatively to improve optimization efficiency. On Qwen3.5-4B math, CoPES recovers 92% of full-parameter GRPO's validation-accuracy gain within the GPU-hour budget of GRPO's best validation checkpoint, versus 67% for standard ES. Its theoretical GPU memory requirement is less than one-eighth that of full-parameter GRPO. It consistently outperforms standard ES and LoRA-based GRPO on all evaluated pass@k metrics across five math benchmarks. CoPES's gains extend to Qwen3.5-9B with unchanged optimization hyperparameters and to multi-hop QA. These results demonstrate an improved trade-off between memory requirements and training time for agentic LLM post-training under resource constraints. The code is open-sourced in https://github.com/MetaronWang/CoPES