- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates whether large language model (LLM) agents can continuously acquire knowledge from tutorial videos without updating model parameters, and proposes Tutorial Agentic Learning (TAL). TAL organizes video content into a structured and continually updated external “tutorialbook” memory. It first generates and filters difficult questions that the model cannot answer independently in order to identify knowledge gaps, and then iteratively refines the tutorialbook based on these failures. At inference time, TAL uses TOC-based retrieval to select relevant chapters for answering queries. The authors introduce VidLearn, a benchmark consisting of 72 chess tutorial videos, and evaluate TAL across multiple mainstream models against Transcript-RAG, Video-RAG, mem0, and ReasoningBank. TAL achieves up to 71.8% accuracy on the held-out chess test set, outperforming the strongest baseline at approximately 64.5%, and also shows cross-domain improvements on Photoshop tutorials. Overall, the paper demonstrates that frozen LLMs can improve downstream task performance by continually constructing and refining structured external memory.
Critical strengths
Strength 1: Important Problem and Coherent Method Design
The paper addresses the practically relevant problem of enabling LLM agents to continuously acquire knowledge from tutorial videos without parameter updates. TAL realizes this through a coherent loop of knowledge-gap identification, structured memory refinement, and inference-time retrieval. In particular, the hard-question harvesting mechanism selectively targets knowledge that the model currently lacks, giving the overall framework a clear motivation and some methodological novelty.
Strength 2: Relatively Comprehensive Experimental Evaluation
The paper compares TAL not only with Transcript-RAG, but also with Video-RAG, mem0, and ReasoningBank, and further includes multi-model evaluation, cross-domain experiments, visual ablations, and retrieval ablations. With 48 videos, full TAL achieves 71.8% held-out accuracy compared with 62.1% for transcript-only TAL, supporting the contribution of visual information. The authors also examine potential train-test leakage. Overall, the experimental evaluation is reasonably comprehensive and provides solid support for the effectiveness of the proposed system.
Critical weaknesses
Weakness 1: Conceptual Gap Between the Learning Framing and the Actual Evidence
The paper frames TAL as enabling LLM agents to “learn by watching tutorial videos,” yet the underlying model remains frozen throughout, and all newly acquired knowledge is stored in an external tutorialbook that remains accessible at test time. Therefore, the current experiments more directly demonstrate that better external memory improves downstream performance, rather than that the model itself acquires persistent new capabilities. Since this issue directly concerns the paper’s central research question and contribution framing, it is my primary concern.
Weakness 2: Lack of a Strong Compute-Matched Structured-Memory Baseline
TAL requires substantial additional LLM computation during memory construction, including question generation, grounding judgment, chapter rewriting, and repeated verification, whereas the main RAG baselines use much lighter preprocessing. Without a compute-matched structured-memory baseline using a comparable LLM-call/token budget, it is difficult to determine whether the gains come from TAL-specific components such as hard-question harvesting, adversarial consolidation, and TOC retrieval, or mainly from the larger preprocessing budget. This weakens the evidence for the claimed superiority of TAL’s specific mechanisms.
Questions for authors
Q1. How do the authors define learning in this paper?
The authors treat the tutorialbook as part of the agent, so updating external memory can be interpreted as nonparametric agent learning. However, since the underlying LLM remains frozen and the performance gains continue to depend on inference-time access to the tutorialbook, I would like the authors to clarify: what is the essential distinction between learning and constructing and consulting an external knowledge resource in this setting? Which experimental results support knowledge acquisition rather than simply improved access to external knowledge?
Q2. To what extent can TAL’s performance gains be explained by additional LLM compute?
TAL requires substantial LLM computation for each video, including question generation, judging, memory rewriting, and verification, whereas the preprocessing used by the main RAG baselines is much lighter. Could the authors provide a compute-matched structured-memory baseline that uses a comparable LLM-call/token budget to summarize, deduplicate, and organize the tutorial videos into structured chapters, but without TAL’s hard-question adversarial loop, while using the same or a similar retrieval mechanism? Such a comparison would help determine whether the gains primarily come from TAL’s proposed mechanisms or from the larger preprocessing budget.
6 comments
Author Rebuttal (Part2)
W2 / Q2. Structured-Memory Baseline Under Comparable Computational Budgets.
Thank you for this valuable suggestion. To examine whether additional LLM computation explains TAL’s performance gains, we introduce an Iterative RAG baseline that constructs structured tutorial chapters from the same video frames and full transcripts through extraction, summarization, and deduplication. Without TAL’s diagnostic QA harvesting, it iteratively refines these chapters by identifying omissions and factual inconsistencies and verifying revisions against the original tutorials.
Both methods use Gemini 3.5 Flash for memory construction and question answering, with the same top-3 TOC retrieval procedure. We evaluate four- and eight-round Iterative RAG checkpoints across 12, 24, 36, and 48 videos on the same 174 training and 110 held-out questions. For Iterative RAG, the training questions are used only for evaluation and do not guide memory refinement. To assess computational comparability, we report both construction-token usage and average per-query inference-token usage, including input and output tokens for retrieval and answer generation.
Table 3: Construction and inference token usage and training/held-out accuracy of TAL and Iterative RAG. Gemini 3.5 Flash backbone; temperature =0.7; accuracy reported as mean±SD over 3 seeds.
| Videos | Method | Construction budget: Total tokens (M) | Construction budget: Ratio to TAL | Inference / query: Avg. total (K) | Inference / query: Ratio to TAL | Accuracy (%) ↑: Train | Accuracy (%) ↑: Test |
|---|---|---|---|---|---|---|---|
| 12 | Iterative RAG (4 rounds) | 4.351 | 0.895× | 7.884 | 1.296× | 60.0±0.7 | 53.9±1.7 |
| 12 | Iterative RAG (8 rounds) | 6.952 | 1.431× | 7.795 | 1.281× | 60.0±1.2 | 54.2±0.9 |
| 12 | TAL | 4.859 | 1.000× | 6.086 | 1.000× | 68.2±1.2 | 64.5±2.0 |
| 24 | Iterative RAG (4 rounds) | 8.820 | 0.878× | 11.110 | 1.154× | 64.8±0.5 | 57.6±1.1 |
| 24 | Iterative RAG (8 rounds) | 14.142 | 1.408× | 11.323 | 1.176× | 65.3±1.8 | 55.2±1.1 |
| 24 | TAL | 10.047 | 1.000× | 9.632 | 1.000× | 70.7±1.7 | 64.8±1.1 |
| 36 | Iterative RAG (4 rounds) | 14.295 | 0.858× | 14.903 | 1.065× | 66.1±0.9 | 60.0±1.3 |
| 36 | Iterative RAG (8 rounds) | 23.143 | 1.390× | 15.031 | 1.074× | 68.6±0.3 | 60.3±1.9 |
| 36 | TAL | 16.655 | 1.000× | 13.990 | 1.000× | 77.8±1.2 | 66.1±0.4 |
| 48 | Iterative RAG (4 rounds) | 18.485 | 0.866× | 17.261 | 0.981× | 64.6±1.4 | 52.4±2.4 |
| 48 | Iterative RAG (8 rounds) | 30.036 | 1.406× | 17.467 | 0.993× | 67.4±1.0 | 56.7±0.9 |
| 48 | TAL | 21.357 | 1.000× | 17.594 | 1.000× | 84.1±1.4 | 71.8±2.0 |
TAL achieves higher mean accuracy than both Iterative RAG variants across all four video budgets on both training and held-out questions. At 48 videos, TAL achieves 71.8% held-out accuracy, compared with 52.4% and 56.7% for four- and eight-round Iterative RAG, respectively. Notably, the eight-round variant uses 1.406× TAL’s construction-token budget with nearly identical per-query inference usage (17.467K vs. 17.594K tokens), yet remains 15.1 percentage points below TAL.
These results suggest that structured knowledge construction and iterative refinement without TAL’s knowledge-gap-driven learning mechanism did not match TAL’s performance in our evaluation, even with greater construction budgets and similar inference usage. We will include this comparison in the revised manuscript.
References.
[1] N. Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS, 2023.
[2] A. Zhao et al. ExpeL: LLM Agents Are Experiential Learners. AAAI, 2024.
[3] S. Ouyang et al. ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory. ICLR, 2026.
[4] P. Asawa et al. Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments. arXiv:2606.05661, 2026.
Dear Reviewer YYVN,
Thank you very much for your time and effort in reviewing our paper. Your constructive comments have been very helpful in improving the quality of our work.
As the author–reviewer discussion period is approaching its end, we would be grateful to hear whether our responses have addressed your concerns or if any questions remain. We would be happy to provide further clarification or additional evidence if needed.
Thank you again for your time and consideration.
Best regards, The Authors
Author Rebuttal (Part1)
Thank you for your constructive feedback and for recognizing the significance of our work and its empirical results. We address your concerns below.
W1 / Q1. Definition of Agent Learning and Supporting Evidence.
Thank you for this valuable question. Our research question concerns the complete LLM agent rather than the underlying LLM alone. In TAL, the agent consists of a frozen LLM and an agent harness that manages persistent memory, retrieval, and context construction.
We define agent learning as acquiring knowledge from instructional materials, retaining it in the agent’s persistent memory, and using it to improve performance on subsequent tasks. Formally, we represent the agent as At=(fθ,Ht), where fθ is the frozen model and Ht is the agent harness, whose persistent memory is instantiated as the tutorialbook Bt in TAL. The model parameters remain fixed, while Bt evolves as the agent learns from instructional materials. This formulation is consistent with prior work on nonparametric agent learning, including Reflexion [1], ExpeL [2], and ReasoningBank [3].
TAL’s distinction from conventional knowledge-base construction and retrieval lies in its learner-dependent diagnostic feedback, which guides what knowledge should be acquired and refined. Specifically, TAL autonomously identifies the agent’s knowledge gaps through tutorial-grounded diagnostic questions (Stage I) and uses these signals to guide memory refinement (Stage II). The acquired knowledge is subsequently reused through TOC retrieval. In this agent-level formulation, inference-time access to the tutorialbook is how the agent utilizes its acquired knowledge, rather than an alternative to learning.
We provide three complementary results supporting this learning formulation.
First, the stage ablation shows that knowledge-gap-guided refinement improves held-out accuracy from 56.36% to 71.82% (+15.46 pp), with both settings using external memory under the same retrieval framework. With the backbone model fixed, accuracy increases from 48.18% to 64.55% as the tutorial corpus grows from 12 to 48 videos, while downstream chess gameplay reduces mean centipawn loss from 162.81 to 125.74. Our structured-memory comparison under comparable computational budgets further examines whether knowledge organization and retrieval alone explain TAL’s gains, as detailed in W2/Q2.
Second, we evaluate cross-opening knowledge generalization through a leave-one-opening-out experiment. For each fold, we construct the tutorialbook using videos and diagnostic questions only from the other eleven openings. We then evaluate only the held-out questions corresponding to the excluded opening, covering all 110 test questions across the twelve folds at each video budget. As the learning corpus grows from 11 to 44 tutorials, accuracy increases from 59.09% to 64.85%, exceeding Vanilla (57.6%) by 7.25 pp at the largest budget. These results provide further evidence that TAL can generalize beyond the specific knowledge covered during memory construction.
Table 1: Leave-one-opening-out evaluation: held-out accuracy (%). Columns indicate tutorials per available opening. Gemini 3.5 Flash backbone; temperature =0.7; mean±SD over 3 seeds.
Third, we further evaluate TAL on Continual Learning Bench (CL Bench) [4], covering all six tasks in the official task environments. In this text-only setting, task descriptions, interaction trajectories, and environment feedback serve as instructional materials, replacing the video-based inputs while preserving TAL’s two-stage knowledge consolidation and TOC retrieval mechanisms. Using Claude Sonnet 4.6 as its backbone, TAL achieves an aggregate normalized reward of 0.3535 and gain of 0.3813 under CL Bench’s official evaluation protocol, ranking first among the 13 compared systems (Table 2) and outperforming the strongest competing system with the same backbone by 0.1304 and 0.1269, respectively.
Table 2: CL Bench results across six tasks (mean ± 95% CI). Baseline results are from the original CL Bench paper [4].
Together, these results support our formulation of nonparametric agent learning. We will clarify this formulation in the revised manuscript.