AI agents rely on a harness of skills, tools, and workflows to solve complex problems. Continually improving this harness is essential for adapting to new tasks. However, existing optimization methods typically require ground-truth validation sets, yet such labeled data is difficult to acquire in practical deployment settings. To address this problem, we introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes the agent harness using only past trajectories. Specifically, RHO selects a diverse coreset of challenging tasks from past trajectories and re-solves them in parallel. The agent analyzes these rollouts using self-validation and self-consistency, then generates candidate harness updates and selects the most effective one by its own pairwise self-preference. We evaluate RHO across three diverse domains, spanning software engineering, technical work, and knowledge work. Notably, a single optimization round improves the pass rate on SWE-Bench Pro from 59% to 78% without any external grading. Furthermore, our analysis demonstrates that RHO effectively targets prior failure modes. As a result, the optimized harness alters the agent’s behavior patterns and sustains higher accuracy during long-horizon sessions. Code is available at https://github.com/wbopan/retro-harness and the project website at https://paper-rho.wenbo.io.

Google’s Agent-to-Agent (A2A) protocol transforms multi-agent collaboration from a closed pipeline into an open and dynamic ecosystem. A2A-based multi-agent systems typically depend on a central orchestrator to coordinate task allocation across agents. However, this paradigm exposes a new attack surface at the agent-orchestration layer: an agent provider can forge an adversarial agent profile to bias the orchestrator’s decisions. We find this risk and propose OrcaJack, an agent orchestration hijacking attack. By injecting a deceptively appealing agent profile into the agent pool, OrcaJack causes user queries in a target domain to be consistently routed to the malicious agent. We formulate the adversarial agent profile construction process as a discrete optimization problem and solve it with a domain-skill-driven genetic algorithm. Across multiple domain benchmarks and orchestrator LLMs, Orcajack can reach up to 100% routing rate, while simple pool-side defenses fail to flag the crafted profiles. Our work is the first to expose the over-trust of agent-declared capabilities in modern agent discovery, motivating pool-side validation defenses.

Data attribution methods often compress the effect of each training data point into a single score, computed either at a final checkpoint or by aggregating quantities along the training trajectory. This static view hides how the contribution of a training data point changes during training. To overcome this limitation, we propose the Window-level Influence Estimator (WIE), a framework for estimating the contribution of a training data point zⱼ over a user-specified training window [t₁, t₂]. WIE models a leave-one-out-style counterfactual under a first-order linearization of SGD. An influence propagator P[t] = I − ηₜH[t] at each step models how parameter deviation is amplified or dampened by the loss landscape. Empirically, the window-level estimator reveals four distinct influence patterns: Early Influencer, Late Bloomer, Stable, Highly Fluctuating. Across multiple benchmarks, WIE improves mislabeled-data detection over the considered static and trajectory-based baselines. On BERT-IMDB with 40% label noise, WIE outperforms these baselines across the evaluated inspection ratios. On ViT-EMNIST, WIE reaches 70% recall for mislabeled data while inspecting up to 7.5× fewer candidates than TracIn requires to reach the same recall.

Large Language Models (LLMs) struggle to automate real-world vulnerability detection due to two key limitations: the heterogeneity of vulnerability patterns undermines the effectiveness of a single unified model, and manual prompt engineering for massive weakness categories is unscalable.To address these challenges, we propose MulVul, a retrieval-augmented multi-agent framework designed for precise and broad-coverage vulnerability detection. MulVul adopts a coarse-to-fine strategy: a Router agent first predicts the top- coarse categories and then forwards the input to specialized Detector agents, which identify the exact vulnerability types. Both agents use evidence retrieved from vulnerability knowledge bases to mitigate hallucinations. Crucially, to automate the generation of specialized prompts, we design Cross-Model Prompt Evolution, a prompt optimization mechanism where a generator LLM iteratively refines candidate prompts while a distinct executor LLM validates their effectiveness. This decoupling mitigates the self-correction bias inherent in single-model optimization. Evaluated on 130 CWE types, MulVul achieves 34.79% Macro-F1, outperforming the best baseline by 41.5%. Ablation studies validate cross-model prompt evolution, which boosts performance by 51.6% over manual prompts by effectively handling diverse vulnerability patterns.

Large language models (LLMs) demonstrate significant capabilities in many natural language processing. However, their performance in machine translation is still behind the models that are specially trained for machine translation with an encoder-decoder architecture. This paper investigates how to improve neural machine translation (NMT) with LLMs. Our proposal is based on an empirical insight that NMT gets worse fluency than human translation. We propose to use LLMs to enhance the fluency of NMT’s generation by integrating a language model at the target side. we use contrastive learning to constrain fluency so that it does not exceed the LLMs. Our experiments on three language pairs show that this method can improve the performance of NMT. Our empirical analysis further demonstrates that this method improves the fluency at the target side. Our experiments also show that some straightforward post-processing methods using LLMs, such as re-ranking and refinement, are not effective.

There exists a discrepancy between the token-level objective during training and the overall sequence-level quality that is expected from the model. This discrepancy leads to issues like exposure bias.To align the model with human expectations, sequence-level objectives are often used to fine-tune pre-trained models.In this paper, we introduce a contrastive preference model that enhances the traditional Plackett-Luce model by incorporating an indicator function. Building upon this novel preference model, we propose Contrastive Preference Learning (CPL), which uses offline samples with list-wise preferences to fine-tune a pre-trained model in Neural Machine Translation. Our experiments, conducted on three language pairs, demonstrate that CPL outperforms not only the vanilla Transformer model but also other token-level and sequence-level baselines. Furthermore, the ablation study highlights the essential role of the proposed indicator function in achieving this improvement.

In Neural Machine Translation, models are often trained with teacher forcing and suffer from exposure bias due to the discrepancy between training and inference. Current token-level solutions, such as scheduled sampling, aim to maximize the model’s capability to recover from errors. Their loss functions have a side effect: a sequence with errors may have a larger probability than the ground truth. The consequence is that the generated sequences may recover too much and deviate from the ground truth. This side effect is verified in our experiments. To address this issue, we propose using token-level contrastive learning to coordinate three training objectives: the usual MLE objective, an objective for recovery from errors, and a new objective to explicitly constrain the recovery in a scope that does not impact the ground truth. Our empirical analysis shows that this method effectively achieves these objectives in training and reduces the frequency with which the third objective is violated. We conduct experiments on three language pairs: German-English, Russian-English, and English-Russian. Results show that our method outperforms the vanilla Transformer and other methods addressing the exposure bias.

Abstractive Text Summarization (ATS) models are commonly trained using large-scale data that is randomly shuffled. However, the impact of data selection and data ordering on ATS models remains a relatively unexplored research area, where a significant challenge lies in accurately assessing the learning difficulty of each training instance. This study introduces a Data Selection Curriculum (DSC) scoring system that incorporates both the difficulty of improving ATS model via an instance and the expected performance on this instance. By selectively excluding excessively simple and overly complex instances, the training efficiency can be optimized. Furthermore, curriculum learning is integrated to accelerate convergence and improve performance by gradually increasing the learning difficulty, inspired by human learners. Experimental results on the CNN/DailyMail dataset demonstrate that our approach surpasses potent baselines, utilizing a mere 20% of the available instances.

Beam search is the most popular decoding method for Neural Machine Translation (NMT) and is still a strong baseline compared with the newly proposed sampling-based methods. To better understand beam search, we investigate its two well-recognized issues, beam search curse and search errors, at the sentence level. We find that only less than 30% of sentences in the test set experience these issues. Meanwhile, there is a related phenomenon. For the majority of sentences, their gold references have lower probabilities than the predictions from beam search. We also test with different levels of model errors including a special test using training samples and models without regularization. We find that these phenomena still exist even for a model with an accuracy of 95% although they are mitigated. These findings show that it is not promising to improve beam search by seeking higher probabilities in searching and further reducing its search errors. The relationship between the quality and the probability of predictions at the sentence level in our results provides useful information to find new ways to improve NMT.