Humans continuously acquire knowledge from carefully crafted materials such as books and tutorial videos, carrying that understanding into future tasks. Millions of such instructional videos are uploaded online every day, yet LLM agents remain unable to exploit this ever-expanding wealth of human teaching. Their capabilities are frozen after training, while parameter updates at scale remain prohibitively expensive. This raises the central question: can LLM agents learn from this vast instructional resource without parameter updates? We introduce Tutorial Agentic Learning (TAL), a nonparametric continual-learning paradigm that consolidates tutorial videos into a persistent, structured tutorialbook memory through a two-stage adversarial process that harvests hard, video-grounded questions and uses them to refine the memory. At inference time, TAL retrieves through a Table-of-Contents (TOC) tree, tying retrieval directly to the model's reasoning rather than to embedding similarity. To study this setting, we introduce VidLearn, a multimodal QA benchmark built from 72 chess tutorial videos spanning 12 openings. On VidLearn, TAL outperforms all evaluated baselines and improves monotonically as more videos are consumed, achieving up to 113% and 91% relative gains over the corresponding vanilla MLLMs on the training and test splits, respectively. In actual gameplay, TAL reaches a stronger opening-phase position than its vanilla counterpart in 69% of head-to-head games. Evaluation in a second domain further demonstrates the generality of TAL. Together, these results highlight instructional videos as a scalable resource for parameter-update-free agent continual learning.
Figures & tables
Figure 1: Overview of learning paradigms for LLM agents, comparing parametric and non-parametric approaches along the dimensions of stored knowledge, inference-time usage, and update mechanism.
Figure 2: Overview of Tutorial Agentic Learning (TAL). In Stage I, TAL performs video-grounded agent–model adversarial consolidation, where a video agent harvests hard question–answer pairs from instructional videos that are supported by tutorial evidence yet unsolved by the frozen model, forming a question bank. In Stage II, TAL performs question–memory adversarial consolidation, using this question bank to iteratively refine tutorial memory through a read–write loop. At inference time, TAL retrieves relevant tutorial units via a Table-of-Contents (TOC) tree, enabling the frozen model to access structured tutorial knowledge for downstream reasoning without parameter updates.
Method
Training Set
Test Set
# Demonstration Videos
12
24
36
48
12
24
36
48
Vanilla MLLMs (Direct Answering)
GPT-5.2
31.61%
31.61%
31.61%
31.61%
37.27%
37.27%
37.27%
37.27%
Kimi K2.5
38.51%
38.51%
38.51%
38.51%
31.82%
31.82%
31.82%
31.82%
Gemini 3 Flash
47.70%
47.70%
47.70%
47.70%
48.18%
48.18%
48.18%
48.18%
Claude Sonnet 4.6
41.95%
41.95%
41.95%
41.95%
37.27%
37.27%
37.27%
37.27%
Table 1: Comparison of baseline methods and TAL on the training and test sets under varying numbers of demonstration videos. Best results are in bold . TAL consistently outperforms all baselines across all settings, showing a clear scaling trend as more tutorial demonstrations become available.
Figure 3: Scaling behavior of TAL vs. vanilla (GPT-5.2) on training (left) and test (right) splits.
Method
Training Set (174 QA)
Test Set (110 QA)
# Demonstration Videos
12
24
36
48
12
24
36
48
Vanilla (no memory)
53.1 ± 1.0
57.6 ± 2.3
Transcript-RAG
55.9 ± 0.5
58.2 ± 1.1
60.0 ± 2.3
64.6 ± 3.1
59.7 ± 1.9
59.4 ± 0.4
59.7 ± 1.5
60.0 ± 2.6
Video-RAG (caption retrieval) [ 38 ]
58.0 ± 0.9
59.6 ± 0.3
59.4 ± 2.7
62.1 ± 1.7
58.5 ± 0.9
57.9 ± 0.9
62.1 ± 1.5
64.5 ± 1.5
Video-RAG (+retrieved frames) [ 38 ]
59.2 ± 2.0
57.3 ± 2.2
60.0 ± 2.7
62.1 ± 2.6
60.3 ± 1.5
61.8 ± 3.7
62.4 ± 2.3
63.6 ± 0.7
mem0 [ 17 ]
57.1 ± 0.5
55.7 ± 1.4
58.2 ± 1.0
59.6 ± 0.7
62.7 ± 2.0
63.0 ± 1.7
62.4 ± 0.4
60.9 ± 1.5
Table 2: Comparison with stronger retrieval and memory baselines. Best results are in bold .
Method
Training Set (53 QA)
Test Set (55 QA)
# Demonstration Videos
4
8
12
16
4
8
12
16
Vanilla (no memory)
38.4 ± 5.4
21.2 ± 3.7
Transcript-RAG
62.3 ± 2.7
66.7 ± 0.9
72.3 ± 0.9
78.6 ± 0.9
27.3 ± 2.6
35.8 ± 0.9
29.7 ± 3.1
30.9 ± 2.6
Video-RAG (caption retrieval) [ 38 ]
68.6 ± 0.9
72.3 ± 2.4
81.8 ± 0.9
84.3 ± 3.2
33.9 ± 0.9
38.2 ± 8.3
29.7 ± 3.1
38.8 ± 4.8
Video-RAG (+retrieved frames) [ 38 ]
67.3 ± 1.8
74.8 ± 3.2
84.3 ± 0.9
86.2 ± 1.8
32.1 ± 0.9
36.4 ± 1.5
27.3 ± 1.5
33.9 ± 2.3
mem0 [ 17 ]
54.1 ± 3.6
57.2 ± 0.9
67.9 ± 1.5
71.7 ± 1.5
29.7 ± 5.6
33.9 ± 1.7
33.9 ± 4.8
33.9 ± 1.7
Table 3: Second-domain evaluation on Adobe Photoshop 2026 software tutorials. Best results are in bold .
Horizon
Win Rate ↑ (TAL/Van)
Mean ACPL ↓ (TAL/Van)
Median ACPL ↓ (TAL/Van)
5% Trimmed Mean ACPL ↓ (TAL/Van)
move2
50.83% / 49.17%
63.14 / 85.26
11.50 / 24.50
47.98 / 74.36
move3
69.17% / 30.83%
87.62 / 142.98
28.83 / 135.17
75.01 / 130.75
move4
67.92% / 32.08%
107.34 / 153.51
66.25 / 146.75
99.96 / 145.38
move5
62.50% / 37.50%
125.74 / 162.81
112.60 / 156.60
119.24 / 159.01
Table 4: Head-to-head TAL vs. vanilla results under fixed openings.
Table 8
Setting
# Demonstration Videos
Top- k
12
24
36
48
k=3
54.55%
57.27%
69.09%
71.82%
k=5
58.18%
60.00%
60.91%
66.36%
Table 8: Impact of retrieval depth.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Opening
The Caro-Kann Defense
The French Defense
The Grünfeld Defense
The Italian Game
The King’s Indian Defense
The London System
The Nimzo-Indian Defense
The Queen’s Gambit
The Ruy Lopez
The Scotch Game
The Sicilian Defense
The Slav Defense
Appendix
Table 9: The 12 classical openings covered in VidLearn.
Figure 4: Interactive user interface used for human evaluation.
Figure 5: Comparison of learning curves on the held-out test set.
Against Train Questions (n=174)
Against Train Subtitle Chunks (n=2564)
Metric
Mean ± SD
95% CI
Median
P90
≥0.8
Mean ± SD
95% CI
Median
P90
≥0.8
Ngram-2 Overlap
0.295 ± 0.116
[0.273, 0.317]
0.282
0.452
0.0%
0.208 ± 0.062
[0.196, 0.219]
0.200
0.286
0.0%
Max Jaccard
0.389 ± 0.073
[0.377, 0.403]
0.393
0.468
0.0%
0.184 ± 0.029
[0.179, 0.190]
0.185
0.219
0.0%
TF-IDF Cosine
0.210 ± 0.065
[0.198, 0.223]
0.207
0.288
0.0%
0.100 ± 0.024
[0.095, 0.104]
0.097
0.131
0.0%
LSA Cosine (128d)
0.419 ± 0.115
[0.397, 0.442]
0.407
0.574
0.0%
0.426 ± 0.079
[0.411, 0.440]
0.423
0.513
0.0%
Appendix
Table 10: Nearest-neighbor overlap between each of the 110 test questions and the 174 training questions (left) / 2,564 training subtitle chunks (right). Lower values indicate weaker train–test similarity. No test question has a near-duplicate match ( ≥0.8 ) under any metric or pool, ruling out trivial lexical leakage.
Adapting large language model agents during deployment requires not only retaining past experience, but also turning new observations into timely guidance. Many test-time learning methods, however, acquire knowledge from completed episodes. Feedback from an ongoing interaction may therefore not be distilled into knowledge soon enough to help the next decision. Acquiring knowledge at the granularity of individual transitions could reduce this delay, but raises a separate challenge: a rule that is useful within one episode may not be reliable enough to guide future episodes. Waiting for validation can forfeit immediate benefits, whereas unrestricted reuse can propagate accidental or misattributed guidance. We introduce StepLearn, a nonparametric framework that separates immediate use from persistent trust. It turns informative transitions into hypotheses that can guide the next step, while requiring prospective validation before reuse across episodes. Their predicted effects are checked against subsequent observations outside the source episodes, and only sufficiently supported hypotheses become available for persistent guidance. This process updates external knowledge while keeping all model parameters fixed. Over five rounds on WebArena-Lite and ALFWorld, StepLearn achieves average success rates of 59.9% and 84.0% with GPT-5-mini, and 57.8% and 88.1% with Qwen3.5-35B-A3B, respectively. It outperforms EvoTest, the strongest baseline, by 2.2-12.7 percentage points across the four settings. Learning dynamics further shows that these gains are not restricted to the final repetition, with advantages already present on first task attempts in most settings.
Deploying LLM agents typically requires a compact test-time student, even if a stronger teacher is available during training. We study knowledge brewing: distilling a teacher's interactive experience into a persistent external memory for the student. Crucially, this requires no weight updates, expert demonstrations, ground-truth labels, or test-time teacher access. This setting poses two challenges: environments provide only sparse, binary feedback, and teacher-authored notes must be inherently tailored to be concretely executable by a substantially weaker student. To address these hurdles, we propose AgentBrew, comprising two coupled components. First, a failure-triggered teacher--Ralph Loop mitigates sparse feedback by transforming student failures into environment-validated notes. Second, student-aware synthesis calibrates teacher knowledge to the weak executor's operational granularity, yielding model-specific, actionable guidance. Extensive evaluations and comprehensive ablations across coding, math, and tool-use tasks demonstrate that this asymmetric, training-free brewing paradigm produces highly capable yet deployable LLM agents.
Video understanding with multimodal large language models (MLLMs) remains challenging due to the long token sequences of videos, which contain extensive temporal dependencies and redundant frames. Existing approaches typically treat MLLMs as passive recognizers, processing entire videos or uniformly sampled frames without adaptive reasoning. Recent agent-based methods introduce external tools, yet still depend on manually designed workflows and perception-first strategies, resulting in inefficiency on long videos. We present EVA, an Efficient Reinforcement Learning framework for End-to-End Video Agent, which enables planning-before-perception through iterative summary-plan-action-reflection reasoning. EVA autonomously decides what to watch, when to watch, and how to watch, achieving query-driven and efficient video understanding. To train such agents, we design a simple yet effective three-stage learning pipeline - comprising supervised fine-tuning (SFT), Kahneman-Tversky Optimization (KTO), and Group Relative Policy Optimization (GRPO) - that bridges supervised imitation and reinforcement learning. We further construct high-quality datasets for each stage, supporting stable and reproducible training. We evaluate EVA on six video understanding benchmarks, demonstrating its comprehensive capabilities. Compared with existing baselines, EVA achieves a substantial improvement of 6-12% over general MLLM baselines and a further 1-3% gain over prior adaptive agent methods.