The Interplay of Harness Design and Post-Training in LLM Agents

Tool-integrated LLM agents are often wrapped within a harness: the scaffolding that determines which tools are exposed, how they are described, and what auxiliary information accompanies each per-step observation. While agents are routinely post-trained, this scaffolding is typically treated as a fixed engineering detail, with design effort limited to the training-free regime. Moreover, existing post-training algorithms assume a static environment, even though tool environments and tasks often shift upon deployment. To address this gap, we extend $\texttt{ALFWorld}$ (i) to treat the harness as a controllable design dimension and (ii) to support evaluation under task and tool environment shifts. Building on this, we systematically analyze how the harness design influences post-training in both in-distribution and out-of-distribution (OOD) settings. We empirically show that harness-aware post-training not only improves in-distribution performance but also enables agents to robustly adapt to OOD settings. Under a harness with minimal design effort, post-training suffers a drastic performance drop under stronger tool environment shifts, further highlighting the importance of harness-aware post-training under such shifts.

Long Live the Librarian! A Persistent Search Sub-Agent for Energy-Efficient Multi-Agent Software Engineering Systems

Multi-agent systems (MAS) have substantially advanced autonomous software engineering (SWE), but their growing inference energy demands raise sustainability concerns. In this paper, we demonstrate that this cost is concentrated in an overlooked source: redundant output tokens generated across agents. Two empirical findings ground this claim. First, our per-token energy attribution for MAS reveals a sharp asymmetry: an output token consumes 30 to 1,000 times more energy than an input or cached token. Second, MAS inflate per-episode output because agents repeatedly re-explore overlapping repository regions. To address this inefficiency, we propose Librarian, a persistent search sub-agent that tracks repository-search history and suppresses redundant exploration actions across agents. By returning short references to file regions instead of full file excerpts, Librarian further reduces output-token volume. On SWE-Bench Verified, Librarian reduces per-episode GPU energy consumption of existing multi-agent SWE systems by up to 25% while preserving task performance.

Concept Unlearning via Cross-Attention Activation Projection for Diffusion Models

Concept unlearning aims to erase a target concept from a pretrained text-to-image diffusion model without retraining. Closed-form methods are attractive in this setting because they apply a single deterministic edit to the cross-attention weights and add no inference-time cost. Existing closed-form methods, however, represent the target concept through the text encoder's response to a few short anchor prompts that name it, and paraphrased prompts that evoke the concept without naming it consistently bypass the edit. We argue that the target should instead be represented in the cross-attention activation space. Text embeddings describe the user's prompt, while cross-attention activations describe what the model is about to render, and the latter generalize to paraphrase the anchor templates do not cover. Building on this observation, we propose PURE (Projection in U-Net Rendering for Erasure), a closed-form method that builds the forget and retain bases from per-layer cross-attention activations captured along a short denoising trajectory and applies a single linear projector to the cross-attention key and value weights. On a recent holistic concept-unlearning benchmark covering ten concepts across artistic style, intellectual property, celebrity, and NSFW categories, PURE significantly reduces target leakage under paraphrased and adversarial prompts while preserving retain concepts close to the unedited model, yielding the best overall forget-retain trade-off among evaluated methods.

Mind the Rarities: Can Rare Skin Diseases Be Reliably Diagnosed via Diagnostic Reasoning?

Large vision-language models (LVLMs) demonstrate strong performance in dermatology; however, evaluating diagnostic reasoning for rare conditions remains largely unexplored. Existing benchmarks focus on common diseases and assess only final accuracy, overlooking the clinical reasoning process, which is critical for complex cases. We address this gap by constructing DermCase, a long-context benchmark derived from peer-reviewed case reports. Our dataset contains 26,030 multi-modal image-text pairs and 6,354 clinically challenging cases, each annotated with comprehensive clinical information and step-by-step reasoning chains. To enable reliable evaluation, we establish DermLIP-based similarity metrics that achieve stronger alignment with dermatologists for assessing differential diagnosis quality. Benchmarking 22 leading LVLMs exposes significant deficiencies across diagnosis accuracy, differential diagnosis, and clinical reasoning. Fine-tuning experiments demonstrate that instruction tuning substantially improves performance while Direct Preference Optimization (DPO) yields minimal gains. Systematic error analysis further reveals critical limitations in current models' reasoning capabilities.

Transductive Generalization via Optimal Transport and Its Application to Graph Node Classification

Many existing transductive bounds rely on classical complexity measures that are computationally intractable and often misaligned with empirical behavior. In this work, we establish new representation-based generalization bounds in a distribution-free transductive setting, where learned representations are dependent, and test features are accessible during training. We derive global and class-wise bounds via optimal transport, expressed in terms of Wasserstein distances between encoded feature distributions. We demonstrate that our bounds are efficiently computable and strongly correlate with empirical generalization in graph node classification, improving upon classical complexity measures. Additionally, our analysis reveals how the GNN aggregation process transforms the representation distributions, inducing a trade-off between intra-class concentration and inter-class separation. This yields depth-dependent characterizations that capture the non-monotonic relationship between depth and generalization error observed in practice. The code is available at https://github.com/ml-postech/Transductive-OT-Gen-Bound.

Training-free Composition of Pre-trained GFlowNets for Multi-Objective Generation

Generative Flow Networks (GFlowNets) learn to sample diverse candidates in proportion to a reward function, making them well-suited for scientific discovery, where exploring multiple promising solutions is crucial. Further extending GFlowNets to multi-objective settings has attracted growing interest since real-world applications often involve multiple, conflicting objectives. However, existing approaches require additional training for each set of objectives, limiting their applicability and incurring substantial computational overhead. We propose a training-free mixing policy that composes pre-trained GFlowNets at inference time, enabling rapid adaptation without finetuning or retraining. Importantly, our framework is flexible, capable of handling diverse reward combinations ranging from linear scalarization to complex non-linear logical operators, which are often handled separately in previous literature. We prove that our method exactly recovers the target distribution for linear scalarization and quantify the approximation quality for nonlinear operators through a distortion factor. Experiments on a synthetic 2D grid and real-world molecule-generation tasks demonstrate that our approach achieves performance comparable to baselines that require additional training.

In-Place Feedback: A New Paradigm for Guiding LLMs in Multi-Turn Reasoning

Large language models (LLMs) are increasingly studied in the context of multi-turn reasoning, where models iteratively refine their outputs based on user-provided feedback. Such settings are crucial for tasks that require complex reasoning, yet existing feedback paradigms often rely on issuing new messages. LLMs struggle to integrate these reliably, leading to inconsistent improvements. In this work, we introduce in-place feedback, a novel interaction paradigm in which users directly edit an LLM's previous response, and the model conditions on this modified response to generate its revision. Empirical evaluations on diverse reasoning-intensive benchmarks reveal that in-place feedback achieves better performance than conventional multi-turn feedback while using $79.1\%$ fewer tokens. Complementary analyses on controlled environments further demonstrate that in-place feedback resolves a core limitation of multi-turn feedback: models often fail to apply feedback precisely to erroneous parts of the response, leaving errors uncorrected and sometimes introducing new mistakes into previously correct content. These findings suggest that in-place feedback offers a more natural and effective mechanism for guiding LLMs in reasoning-intensive tasks.

Delving into Instance-Dependent Label Noise in Graph Data: A Comprehensive Study and Benchmark

Graph Neural Networks (GNNs) have achieved state-of-the-art performance in node classification tasks but struggle with label noise in real-world data. Existing studies on graph learning with label noise commonly rely on class-dependent label noise, overlooking the complexities of instance-dependent noise and falling short of capturing real-world corruption patterns. We introduce BeGIN (Benchmarking for Graphs with Instance-dependent Noise), a new benchmark that provides realistic graph datasets with various noise types and comprehensively evaluates noise-handling strategies across GNN architectures, noisy label detection, and noise-robust learning. To simulate instance-dependent corruptions, BeGIN introduces algorithmic methods and LLM-based simulations. Our experiments reveal the challenges of instance-dependent noise, particularly LLM-based corruption, and underscore the importance of node-specific parameterization to enhance GNN robustness. By comprehensively evaluating noise-handling strategies, BeGIN provides insights into their effectiveness, efficiency, and key performance factors. We expect that BeGIN will serve as a valuable resource for advancing research on label noise in graphs and fostering the development of robust GNN training methods. The code is available at https://github.com/kimsu55/BeGIN.

Influence Functions for Edge Edits in Non-Convex Graph Neural Networks

Understanding how individual edges influence the behavior of graph neural networks (GNNs) is essential for improving their interpretability and robustness. Graph influence functions have emerged as promising tools to efficiently estimate the effects of edge deletions without retraining. However, existing influence prediction methods rely on strict convexity assumptions, exclusively consider the influence of edge deletions while disregarding edge insertions, and fail to capture changes in message propagation caused by these modifications. In this work, we propose a proximal Bregman response function specifically tailored for GNNs, relaxing the convexity requirement and enabling accurate influence prediction for standard neural network architectures. Furthermore, our method explicitly accounts for message propagation effects and extends influence prediction to both edge deletions and insertions in a principled way. Experiments with real-world datasets demonstrate accurate influence predictions for different characteristics of GNNs. We further demonstrate that the influence function is versatile in applications such as graph rewiring and adversarial attacks.

The Oversmoothing Fallacy: A Misguided Narrative in GNN Research

Oversmoothing has been recognized as a main obstacle to building deep Graph Neural Networks (GNNs), limiting the performance. This position paper argues that the influence of oversmoothing has been overstated and advocates for a further exploration of deep GNN architectures. Given the three core operations of GNNs, aggregation, linear transformation, and non-linear activation, we show that prior studies have mistakenly confused oversmoothing with the vanishing gradient, caused by transformation and activation rather than aggregation. Our finding challenges prior beliefs about oversmoothing being unique to GNNs. Furthermore, we demonstrate that classical solutions such as skip connections and normalization enable the successful stacking of deep GNN layers without performance degradation. Our results clarify misconceptions about oversmoothing and shed new light on the potential of deep GNNs.