A comprehensive re-evaluation of two memory-based methods for self-improving agents is conducted, broadening the scope of evaluation along two axes and hypothesizing that task and environment underspecification contribute to this fragility.
Abstract
Memory-based self-improving agents--those that learn from an online stream of tasks and improve over time by maintaining a textual memory bank--have shown great promise in recent literature. However, the reliability aspects of these methods have been critically overlooked. In this work, we conduct a comprehensive re-evaluation of two memory-based methods, broadening the scope of evaluation along two axes: (1) including multiple runs to quantify variance, and (2) randomly shuffling the tasks to investigate the effect of task order. Through these experiments, we make two observations that expose the fragility of current methods: First, agent evaluation is inherently noisy in complex environments and on multi-step tasks, and stacking a self-improving loop on top can further amplify this noise. Second, the agent's improvement is highly dependent on task order. Prior works often adopt default orderings that impose an implicit curriculum, acting as a hidden prerequisite for success. To better understand this fragility, we manually examine the agents'memory and hypothesize that task and environment underspecification contribute to this fragility. We validate this hypothesis by incorporating information that enables better specification, such as detailed rubrics and environment feedback, into the memory construction process. While this added information partially closes the performance degradation in previous experiments, significant gaps still remain, suggesting that other uncharacterized factors contribute to this fragility. Looking ahead, our work advocates for more rigorous evaluation protocols for self-improving agents by reporting results across multiple runs and stress-testing them under challenging conditions. Moreover, our findings on underspecification call for systems and interfaces that enable effective human oversight, preventing agents from failing in unforeseeable ways.
Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain preferences, task histories, tool routines, and learned skills across sessions. Yet whether retained experience actually improves them over time has not been systematically tested. We introduce PAST-Bench, a benchmark designed to isolate this question. Each agent runs through ordered sequences of fresh-session tasks under matched conditions that turn retained experience on and off. It spans 26 scenarios and 204 episodes across memory, procedural reuse, information gathering, and update. We report both later-task gains and whether those gains follow the intended save, retrieve, and update pathway. Across seven base models and four agent frameworks, improvement is real but uneven across capabilities. Agents with the same headline gain can differ markedly in whether that gain is supported by evidence of the intended pathway. Guided by these findings, we develop Hermes+, which extends Hermes with five targeted interventions across stages of the agent loop. Hermes+ raises the average gain from retained experience and provides clearer pathway evidence, with its strongest improvement on tasks requiring outdated state to be replaced, although the effect remains capability- and model-dependent. Together, PAST-Bench and Hermes+ provide an evaluation and diagnostic foundation for studying how persistent agents can progress from retaining experience to systematically improving through it. Code: https://github.com/Gen-Verse/PAST-Bench
Shuhan Xue, Zixin Ding, Yi-Jun Shen et al.· 0 citations
Search agents enables large language models (LLMs) to iteratively interleave retrieval and reasoning, yielding strong performance on knowledge-intensive tasks. However, their multi-step autonomy also introduces substantial inefficiencies. In practice, search agents often exhibit overretrieval, where redundant or irrelevant documents are repeatedly fetched, and overthinking, where reasoning steps become excessive or unproductive. Both behaviors significantly inflate retrieval and inference cost, yet remain poorly understood, particularly under model scaling. In this work, we conduct a systematic study of overthinking and overretrieval in search agents from a scaling perspective. We formalize both phenomena at the trajectory level and propose fine-grained evaluation protocols that combine automatic statistics with LLM-based judgments. Through controlled experiments across search agents built on LLMs of varying sizes, we find that increasing model capacity generally alleviates both behaviors, but to markedly different extents. Building on these analysis results, we further propose a lightweight post-hoc reflection framework that converts the proposed evaluation signals into explicit feedback rewards to guide agents' reasoning trajectories. Our findings provide a principled foundation for diagnosing and controlling inefficiencies in search agents.
Xin Liu, Ruqing Zhang, Yu-An Liu et al.· Annual International ACM SIG...· 0 citations
Self-distillation (SD) has emerged as a compute-efficient alternative to reinforcement learning with verifiable rewards: a self-teacher, conditioned on privileged information (PI) about the answer such as a reference solution, supplies dense per-token supervision to a student that never sees it. Reported gains, however, come almost exclusively from narrow, low-difficulty settings, leaving open a basic question: as a lone objective, with no reward term, does SD teach anything? We reproduce SDPO's reported gains in its easy setting, then apply the identical setup to difficult tasks and find that it does not. Across question answering, mathematics, coding, and multi-turn agentic tool use, across reasoning modes, model sizes, and forms of PI, and under both the SDPO and OPSD recipes, the per-token loss falls steadily while validation accuracy does not improve and typically degrades. We explain this failure through a single causal chain from the loss to the model it produces. The chain begins with PI bias: having seen one particular reference solution, the teacher's per-token target is pulled toward that trajectory rather than toward correctness in general, an effect we quantify with a PI Bias Score. Trained to match this target everywhere, the student's objective becomes nearly blind to whether a rollout is correct, and the loss it assigns falls mostly on low-information tokens like stopwords, punctuation, uncertainty markers, rather than those that determine the answer; within correct rollouts the exploratory tokens incur the highest divergence, so it penalizes the hesitation that reasoning requires. The result is a flatter, less decisive student that is no better at reasoning: as a lone objective, SD optimizes a signal decoupled from task success.
Sarthak Harne, Chinmay Karkar, Yash Pandya et al.· 1 citation
Web search agents powered by Large Language Models (LLMs) show strong promise, but deep research tasks expose a recurring failure mode: once an agent has produced a query, plan, or intermediate conclusion, it becomes less objective when later judging the consequences of that same action. We term this phenomenon inertia bias. To make it measurable, we introduce the IBIS benchmark, which controls the search observations while varying whether the model is evaluating the outcome of its own prior action. We find that models are substantially worse when they"own"the preceding search step, showing that self-authored action history can systematically distort subsequent judgment. We further show that this bias propagates into two forms of system-level degradation: search noise at the worker level and contextual noise at the manager level. To address this problem, we propose NIS-Agent, which applies context isolation at the two decision points most vulnerable to inertia bias: webpage triage and final-answer validation. Across GAIA, WebWalkerQA, BrowseComp, and BrowseComp-zh, NIS-Agent achieves competitive performance while reducing token cost by 33% compared to our baseline. We further train an 8B model to be intrinsically more resistant to inertia bias; under the same NIS-Agent framework, it attains average performance comparable to GPT-4o on deep research benchmarks. Our code is publicly available at https://github.com/PangSMPang/NIS-Agent.
Xiangdong Zhang, Zhanwei Zhang, Zhihang Fu et al.· 0 citations
Skills have emerged as a practical and effective approach for enhancing LLM agents at inference time through structured packages of knowledge. However, existing evaluations largely measure whether skills improve aggregated task success, leaving a more fundamental question underexplored: \emph{\textbf{When do skills help, why do they work, and where do they fail?}} Through controlled experiments across various benchmarks, agent harnesses and LLMs, we isolate the effects of representation, outcome annotation, retrieval difficulty, and cross-framework robustness of skills. To further answer this question, we design a contrastive study that combines controlled quantitative experiments with paired trajectory analysis. We normalize 8,135 trial records from controlled experiments and retain 238 valid unique labels from 240 open-coded records. We consolidate these observations into a taxonomy of three high-level categories and twelve skill-use modes: skills work when noisy trajectories become procedural anchors that stabilize execution. Skills improve over Workflow Memory by 6.06 points in matched comparisons. Procedural anchoring accounts for 65.7\% of skill cases, versus 4.5\% for explicit knowledge injection, showing that skills stabilize action rather than inject missing facts. Retrieval is a separate bottleneck: as pools grow from 5 to 100, actual-use precision falls from 29.6\% to 3.3\%. Confusable distractors impair offline identification, yet downstream success remains stable; exact ground-truth invocation is neither sufficient nor necessary. Skills fail under brittle assumptions, incompatible contexts, or insufficient adaptation. These findings move evaluation beyond aggregate success rates and guide reliable self-evolving agents.
Zhiyuan Jiang, Fan Huang, Hanwen Xing et al.· 0 citations
Lifelong LLM agents increasingly adapt through external learning states that store past interactions as retrievable memories or reusable skills, yet existing benchmarks rarely account for how the path of accumulated experience shapes what agents transfer and retain. In this work, we establish PATH-Bench, a benchmark for path-dependent evaluation of lifelong agents. PATH-Bench estimates directed task relationships via multi-model in-context learning, constructs probe-centered sequences with controlled helpful and interfering histories, and repeatedly evaluates probe tasks to measure average performance, forward transfer, backward transfer, and forgetting. We evaluate eight representative agents on single-turn code generation and multi-turn tool-use tasks under positive- and negative-dominant histories. Benchmark results show that experience utility depends jointly on how experience is represented and on the task's interaction structure, that strong transfer does not ensure retention, and that later experience can reshape gains acquired earlier in the learning path. Based on these findings, we propose Selective Experience Use (SEU), an agent harness that regulates how path-accumulated experience influences each new task, admitting helpful items while filtering out potential interference. SEU consistently reduces forgetting while improving forward transfer in the majority of settings. The PATH-Bench provides both a controlled evaluation framework and actionable guidance for designing more selective and robust lifelong agents.
Xidong Yang, Xingyi Zhang, Wenhao Li et al.· 0 citations
