Masking Stale Observations Helps Search Agents -- Until It Doesn't: A Regime Map and Its Mechanism

• URL: http://arxiv.org/abs/2606.00408v1
• Date: Fri, 29 May 2026 22:51:58 GMT
• Title: Masking Stale Observations Helps Search Agents -- Until It Doesn't: A Regime Map and Its Mechanism
• Abstract summary: Long-horizon search agents accumulate large amounts of retrieved content across many tool calls.<n>We study observation masking through a systematic sweep over various agent backbones and three retrievers on offline and live-web agentic search benchmarks.<n>We find that the accuracy gain from masking follows an asymmetric inverted-U shape when plotted against the model's accuracy without context management.
• Score: 23.342723567214733
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Long-horizon search agents accumulate large amounts of retrieved content across many tool calls, making context-budget efficiency increasingly important. A minimal intervention is to mask stale observations from the context as the trajectory progresses, but it remains unclear when this form of context management helps and why. We study observation masking through a systematic sweep over various agent backbones (4B to 284B parameters) and three retrievers on offline and live-web agentic search benchmarks. We find that the accuracy gain from masking follows an asymmetric inverted-U shape when plotted against the model's accuracy without context management: a plateau under weak retrievers, a peak when a strong retriever meets a mid-capacity model, and a sharp collapse when the model is saturated. This pattern reflects the interaction between retriever recall and the model's implicit filtering capacity, rather than either factor in isolation. Mechanistically, masking implements a token-for-turn trade-off: it removes observations the model has largely stopped attending to and pages the agent rarely re-opens. The added turns help when they convert failures into successes, but they fail when masking removes evidence the model would otherwise have used. We therefore reframe context management as a regime-dependent intervention and provide a holistic perspective for analyzing context use in agentic deep search. We release our scaffold and trajectories here (https://github.com/i-DeepSearch/observation-masking) to support future research.

Related papers

• MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents [74.83535434786145]
  MNIST-PRO is a benchmark that isolates agentic perception by converting MNIST digit recognition into a sequential, glimpse-based search task with lookback constraints.<n>We evaluate ten multimodal models across four memory representations, including raw visual history, textual states, structured metric grid maps, and a consolidated visual canvas.
  arXiv  Detail & Related papers  (2026-08-31T16:05:39Z)
• WeAgent-MMSearch: Native Text-Vision Interaction for Multimodal Search Agents [53.426679066400844]
  We introduce WeAgent-Harness, a multimodal agentic harness that supports native text-vision interaction and runtime recovery.<n>We develop WeAgent-MMSearch, an integrated system spanning data construction, agentic post-training, and multimodal rollout.<n>We also introduce VisTarget-Bench, a 150-task human-verified benchmark that pairs each question with a held-out target image.
  arXiv  Detail & Related papers  (2026-08-28T08:28:43Z)
• Buried in Textual Debt: Context Pruning with Visual Evidence Preservation for MLLM Agents [16.82888862043366]
  We propose a Kullback-Leibler-guided framework for pruning accumulated reasoning in tool-use agents.<n> SPARE achieves the highest average accuracy among pruning methods while removing 37.89-64.58% of reasoning tokens.<n>This favorable accuracy-context trade-off shows that reducing textual dominance restores reliance on visual evidence and mitigates over-conditioning on self-generated language.
  arXiv  Detail & Related papers  (2026-08-24T08:29:17Z)
• Do Agents Think Deeper? A Mechanistic Investigation of Layer-Wise Dynamics in Sequential Planning [6.5668779922726666]
  Large language models (LLMs) may utilize their depth inefficiently in standard single-turn tasks.<n>We show that agentic reasoning exhibits a distinct depth profile from static tasks.
  arXiv  Detail & Related papers  (2026-05-27T04:08:08Z)
• RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations [51.43574078961796]
  We introduce RepoMirage, a two-stage evaluation suite built on SWE-Bench Verified.<n>RepoMirage-Perturb applies three types of semantics-preserving repository-level perturbations, revealing a clear performance drop when correct solving requires broader context access.<n>We propose RepoAnchor, a structure-first prototype workflow that separates repository exploration from downstream problem solving.
  arXiv  Detail & Related papers  (2026-05-25T06:26:43Z)
• Detecting Semantic Backdoors in a Mystery Shopping Scenario [0.0]
  We tackle the problem of detecting semantic backdoors in classification models.<n>Under the assumption that the clean training dataset and the training recipe of the model are both known, we propose a reference model pool.<n>We experimentally analyze a number of approaches to compute model distances and we also test a scenario where the provider performs an adaptive attack to avoid detection.
  arXiv  Detail & Related papers  (2026-01-07T11:04:04Z)
• Backdoor Cleaning without External Guidance in MLLM Fine-tuning [76.82121084745785]
  Believe Your Eyes (BYE) is a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples.<n>It achieves near-zero attack success rates while maintaining clean-task performance.
  arXiv  Detail & Related papers  (2025-05-22T17:11:58Z)
• Towards Competitive Search Relevance For Inference-Free Learned Sparse Retrievers [7.976154147999298]
  inference-free sparse models lag far behind in terms of search relevance when compared to both sparse and dense siamese models.<n>We propose two different approaches for performance improvement.<n>First, we propose an IDF-aware penalty for the matching function that suppresses the contribution of low-IDF tokens.<n>Second, we propose a heterogeneous ensemble knowledge distillation framework that combines siamese dense and sparse retrievers.
  arXiv  Detail & Related papers  (2024-11-07T03:46:43Z)
• Boot and Switch: Alternating Distillation for Zero-Shot Dense Retrieval [50.47192086219752]
  $texttABEL$ is a simple but effective unsupervised method to enhance passage retrieval in zero-shot settings. By either fine-tuning $texttABEL$ on labelled data or integrating it with existing supervised dense retrievers, we achieve state-of-the-art results.
  arXiv  Detail & Related papers  (2023-11-27T06:22:57Z)
• Spatial-Temporal Graph Enhanced DETR Towards Multi-Frame 3D Object Detection [54.041049052843604]
  We present STEMD, a novel end-to-end framework that enhances the DETR-like paradigm for multi-frame 3D object detection. First, to model the inter-object spatial interaction and complex temporal dependencies, we introduce the spatial-temporal graph attention network. Finally, it poses a challenge for the network to distinguish between the positive query and other highly similar queries that are not the best match.
  arXiv  Detail & Related papers  (2023-07-01T13:53:14Z)
• Mask and Restore: Blind Backdoor Defense at Test Time with Masked Autoencoder [50.1394620328318]
  Existing backdoor defense methods often require accessing a few validation data and model parameters.<n>We propose blind backdoor defense with Masked AutoEncoder (BDMAE)<n>BDMAE detects possible local triggers using image structural similarity and label consistency between the test image and MAE restorations.
  arXiv  Detail & Related papers  (2023-03-27T19:23:33Z)
• Improving Distantly Supervised Relation Extraction with Self-Ensemble Noise Filtering [17.45521023572853]
  We propose a self-ensemble filtering mechanism to filter out noisy samples during the training process. Our experiments with multiple state-of-the-art relation extraction models show that our proposed filtering mechanism improves the robustness of the models and increases their F1 scores.
  arXiv  Detail & Related papers  (2021-08-22T11:23:36Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.