acceptodds
Under review as a conference paper at ICLR 2027

AgriSEED: Learning to Teach Agricultural Vision-Language Models

Abstract

Generating useful supervision is a central challenge in adapting vision–language models to agricultural tasks. Although models can generate plausible question–answer pairs, answer validity alone does not indicate whether an example will help a solver learn. We introduce AGRISEED (Solver-Evaluated Example Design), a learning-to-teach framework that optimizes question generation using feedback from temporary solver adaptation. Given agricultural images, a questioner proposes question–answer pairs, and a frozen visual judge screens their quality. For each reliable candidate, a temporary solver undergoes a single low-rank adaptation step. Its learning gain is measured through changes in reference-answer likelihood on screened, self-generated probes from other images, complemented by augmented views of the candidate image. The temporary update is reset after evaluation, and the learning gain is combined with format and answer-quality rewards to train the questioner using GRPO. The trained questioner then generates a filtered dataset for a subsequent stage of solver reinforcement learning under the fixed visual judge. Experiments on AgroCoT with Qwen2.5-VL-7B, Qwen3-VL-8B, and InternVL2-8B show consistent improvements in semantic similarity and chainof-thought quality across all five agricultural task dimensions. A task-wise analysis shows consistent improvements in both semantic alignment and reasoning-step matching across the evaluated agricultural tasks and model backbones, supporting the effectiveness of AGRISEED for agricultural vision–language understanding.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agricultural visual question answering, reinenforcement learning, self-envolving

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_oexhOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes AgriSEED, a learning-to-teach framework that generates the most effective question-answer pairs by leveraging real-world learning gains and visual quality assessments to increase the VLM's learning efficiency in agriculture.

Critical strengths
  1. The authors propose a new framework to solve the adaptation challenge for agricultural tasks.
  2. The models trained via AgriSEED outperformed the baseline models, demonstrating the effectiveness of the framework.
Critical weaknesses
  1. I think the novelty of AgriSEED is limited. This is because existing research has already explored approaches to fine-tuning teacher models or data generators based on the learning outcomes of student models (i.e., learning-to-teach framework). While the authors highlight the framework as a key contribution (I believe), it appears that they have not fully articulated a clear distinction from existing methodologies. I think the main contribution of this study is not in proposing a new learning framework, but in applying the existing learning-to-teach approach to the agricultural domain and experimentally verifying its effectiveness.
  2. To emphasize the authors' proposed framework as a core contribution, a more depth and comprehensive analysis of the roles and effects of each component is required. To gain a deeper understanding, it's essential to systematically investigate how the cross-image probe specifically enhances learning performance, how its performance varies depending on the question-answer generation strategy, and how the design of the reward function impacts both the learning process and final performance. The current paper primarily relies on an ablation study, which limits the analysis to only partially explaining the effectiveness of the proposed framework and the validity of its design choices.
Questions for authors
  1. The scores used in equation (1) is unclear. The authors need an explanation of whether the score was empirically determined or etc. This also applies to the scores used in other formulas. To validate the proposed scoring design, it's essential to conduct additional experiments on how the model's performance changes when different scoring methods or weights are applied.
  2. This study utilizes the AgroCoT dataset for both training and evaluation. Even with only 10% of the total data used in the learning process, evaluating the proposed method solely within the same dataset is insufficient to fully validate its generalization performance. To validate the proposed framework, I recommend further evaluation on external datasets like AgroBench.
  3. Does the Qwen3.5-27B judge align with human judgement?
  4. In Table 3, 10% target => 20% target..? But, the authors mention they used only 10% of AgroCoT. What happened?
  5. As I mentioned above, there's a typo in the sampling count (I think).
2 comments

Teaching signal and probe design. AgriSEED trains the questioner using the effect of a single generated QA pair on the solver. We apply one temporary LoRA step to the solver, compare reference-answer log-likelihood on the same probes before and after the update, and reset the adapter after each candidate. Same-image augmentations measure how the gain carries across views, while cross-image probes measure how it carries to other visual content. At a fixed probe count, the probe-composition analysis can compare each probe type and their combination.

Reward scoring and sensitivity. We will specify the rubric and score ranges in Eq. (1), the reward weights and thresholds, and which settings were fixed or tuned. In the matched comparison on held-out AgroCoT images, a Qwen2.5-VL-7B solver gains +4.2 SS and +3.2 CQ over the base with the original difficulty reward and +5.4 SS and +4.2 CQ with learning gain. This is about 30% more gain relative to the original reward's improvement over the base.

Image-level generalization. We use the 10%-target split for adaptation and reserve the remaining 90% of AgroCoT for held-out testing. We group all questions and augmented views by original image and exclude the held-out images from all adaptation stages, including probe feedback. AgriSEED takes only images from the AgroCoT adaptation split. It generates the training QA pairs and uses feedback from a frozen visual judge. AgriSEED's held-out scores are close to the full-set scores reported in the paper, supporting generalization to unseen images within AgroCoT.

External evaluation on AgroBench. We use the full AgroBench benchmark with five-option exact matching and report macro-averaged accuracy across its seven tasks. The AgriSEED score is 71.3%. The published full-set scores are 70.66% for QwenVLM-72B, 72.24% for Gemini 1.5 Pro, and 73.45% for GPT-4o (Shinoda et al., ICCV 2025, Table 2).

Sampling targets. The 10% and 20% sampling targets came from different experimental versions. All results in the paper use the 10% target. We will correct the sampling entries in Table 3 and Appendix A in the revision to reflect that version.

Human judge agreement. We randomly sampled 10% of the audit pool, which consisted of QA pairs generated during training. Human reviewers and the frozen Qwen3.5-27B judge agreed on 92% of the reviewed pairs.

Thanks for the author's response. However, after review the response, I can't find any evidence to increase the score. Especially, my concern regarding Weakness 2 has not been fully addressed (other concerns also remain). I highly recommend showing more analysis of why the proposed framework is a core contribution. Therefore, I maintain my score.

Reviewer_QLmgOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the problem in agricultural visual-language models (VLMs) where "question-answer pairs are grammatically and factually effective but not necessarily pedagogically valuable for the solver" during synthetic supervised signal adaptation. To solve this problem, the AGRISEED (Solver-Evaluated Example Design) framework is proposed. Inspired by the concept of "learning-to-teach," this method divides the training process into two phases that alternate across multiple iterations:

  • Phase 1 (Question Generator Training): Given an unlabeled agricultural image, the question generator QϕQ_\phi samples and generates open-ended question-answer pairs e=(x,q,a^)e=(x, q, \hat{a}). These pairs are first scored and gated for correctness, completeness, visual validity, and relevance by a frozen visual evaluator (Qwen3.5-27B). For reliable candidate samples that pass the screening, the framework attaches a lightweight LoRA adapter to the frozen base solver Sθ0S_{\theta_0} and performs a one-step temporary SGD update with gradient clipping (the adapter state u0u_0 is reset after each evaluation). Subsequently, the question generator is optimized using the GRPO algorithm by calculating the median of the self-generated cross-image probe set Pe\mathcal{P}_e and the length-normalized log-likelihood gain Δe\Delta_e on the same image augmentation view before and after the update, combined with the format score and the evaluator quality score.

  • Second stage (final solver training): The trained question generator is frozen to generate a candidate question-answer pool. This pool is then filtered using visual evaluator validation and initial solver difficulty sampling (retaining samples in the medium difficulty range of 0.30≤d(e)≤0.700.30 \le d(e) \le 0.70). Finally, the solver is trained using GRPO with the evaluator's answer quality reward.

The authors conducted three iterative experiments on five agricultural dimensions of the AgroCoT benchmark (Quantitative Analysis (QA), Spatial Understanding (SU), Object Detection (OD), Disease Monitoring (DM), and Environmental Management (EM), based on three backbone models: Qwen2.5-VL-7B, Qwen3-VL-8B, and InternVL2-8B. Experimental results show that AGRISEED improves the semantic similarity (SS) and thought quality (CQ) scores by an average of 4.75 and 4.08 points respectively compared to the base model. Its overall performance is significantly better than the adapted game-theoretic baseline SPICE, and it surpasses the supervised GRPO model using real-labeled data on both backbone models.

Critical strengths
  • Addressing the core pain points of self-game theory and synthetic data generation with a clear motivation: The paper accurately points out the limitations of current multimodal self-game methods (such as SPICE, R-Zero, and VisPlay) that use "the variance or uncertainty of the current solver's answer" as a reward—in agricultural scenarios, high uncertainty often stems from ambiguity in disease symptoms or insufficient visual evidence in images, rather than the high educational value of the samples themselves. By directly measuring the cross-image and enhanced view generalization gains after a one-step temporary adaptation, "problem difficulty" is transformed into "actual learning gain," and the conceptual design is highly inspiring.

  • The reward coupling design using second-order derivatives takes into account engineering feasibility: Traditional meta-learning sample reweighting or dataset distillation usually requires calculating the second-order gradient of the inner learner's update process, which is computationally too costly on large-scale vision-language models. This paper obtains the scalar learning gain Δe\Delta_e through a reset-style single-step LoRA update, and then directly uses it as a truncated scalar reward training generator for reinforcement learning (GRPO). This cleverly avoids the backpropagation differentiation of the solver update process. In Section 3.5 and Appendix B, a local theoretical explanation of gradient direction alignment is given using first-order Taylor expansion and descent lemma.

Critical weaknesses
  • The transductive setup compromises the rigor of the experimental results:

    • In Section 4.1 and Appendix A, the authors state that a subset is extracted from AgroCoT as the pool of adapted images, but ultimately the evaluation is performed on all 4,759 samples (including the adapted subset itself). More seriously, the baseline method, “Supervised GRPO,” is trained directly on the adapted subset containing real question answers and CoT labels, and then tested on the full set of benchmarks containing that training subset.

    • This practice of mixing the training set with the test set confounds the evaluation results on “training set memory/overfitting effects” and “true generalization ability.” The authors do not separately report performance on a completely disjoint hold-out test split, nor do they test zero-shot transfer ability on external benchmarks such as AgroBench, AgMMU, or PlantExpertVQA mentioned in related work.

  • The knowledge distillation effect and computational cost of the 27B visual evaluator are not fully discussed:

    • The experiment used Qwen3.5-27B as the visual evaluator JJ for all 7B/8B backbone models. This was not only used to select reference answers for the question generator but also directly provided answer quality rewards RS(y;e)=rJ(x,q,y;a^)R_S(y; e) = r_J(x, q, y; \hat{a}) for the GRPO training of the final solver. This means that AGRISEED essentially involves a process of preference and knowledge distillation from the large 27B model to the small 7B/8B models, while the base model did not benefit from the signal from this strong external model.

    • In the ablation experiments shown in Table 2, "removing the learning gain reward" achieved a score of 79.71 (SS) / 37.76 (CQ) solely based on the judge's quality reward and selection, representing a significant improvement over the base model (76.61 / 35.05). The addition of the core "learning gain reward" only further improved the score by 2.22 / 1.54 points. Furthermore, each reliable candidate sample requires a single-step LoRA backpropagation and multiple forward probe inferences (requiring an additional two A100 80GB GPUs), but the paper makes no report on the actual training time (wall-clock time) and computational cost compared to ordinary GRPO or SPICE.

Questions for authors
  • What is the true generalization performance on the held-out test set? Please resubmit the supervised GRPO, SPICE, and AGRISEED (Iter 1–3) SS and CQ metrics on the remaining AgroCoT samples (image-disjoint test set) after removing all visible images from the adaptation period. Supplementing the paper with zero-shot evaluations across datasets on external benchmarks such as AgroBench or AgMMU would significantly enhance its persuasiveness.
  • How does the self-judgment ablation of the same size perform? If the external 27B strong model is removed and a frozen base model of the same size as the solver is used instead (e.g., the frozen Qwen2.5-VL-7B itself is used as the visual judge JJ) to perform quality screening and reward scoring, how much of the performance improvement of AGRISEED compared to the base model is retained?
2 comments

Generalization on unseen images. The paper's full-set evaluation is transductive. Its results use the 10% adaptation target. We will correct the 20% entries from a different experimental version in Table 3 and Appendix A. For the held-out evaluation, all methods adapt on the 10% split and test on the remaining 90%. We group questions and views by original image and keep held-out images out of adaptation, including probes. Supervised GRPO uses gold QA and CoT on the adaptation split. AgriSEED takes images from AgroCoT and trains on generated QA pairs with frozen-judge feedback. Its held-out performance is close to the results reported in the paper.

Adaptation and held-out scores for Qwen2.5-VL-7B

MethodSeen SSSeen CQHeld-out SSHeld-out CQ
Base76.735.276.535.0
Supervised GRPO81.639.280.638.4
Adapted SPICE79.737.579.137.0
AgriSEED Iteration 180.538.580.338.3
AgriSEED Iteration 281.239.081.038.8
AgriSEED Iteration 382.239.581.939.2

The seen columns describe the 10% adaptation subset. The held-out columns describe the 90% test subset.

Contribution of the visual judge. The frozen Qwen3.5-27B screens generated pairs and scores solver rewards. Human reviewers and the judge agreed on 92% of a random 10% audit sample of generated QA pairs. In Table 2, learning gain adds 2.22 SS and 1.54 CQ points over the no-learning-gain control. In the held-out comparison with the judge fixed, the gains over base increase from +4.2/+3.2 SS/CQ with the difficulty reward to +5.4/+4.2 with learning gain.

Training cost. Visual judging and temporary LoRA updates increase total wall-clock training time by about 40% relative to the configuration before adding these components. Each accepted candidate adds one LoRA update and pre- and post-update probe evaluations. We consider the cost worthwhile given the performance gains when adapting from unlabeled images.

Same-size self-judge

The self-play-style control uses a frozen copy of the same Qwen2.5-VL-7B initialization as the trainable solver. The frozen copy selects reference answers, screens generated QA pairs, and scores solver rewards. The comparison keeps the data split, evaluation, reward rubric, and training budgets fixed. We compare both reward conditions under each judge.

AgroCoT self-judge comparison in SS and CQ points

Condition27B judge (Table 2)7B self-judge
Frozen solver reference76.61 / 35.0576.61 / 35.05
No learning gain79.71 / 37.7679.71 / 37.76
Success-rate reward—80.70 / 38.20
AgriSEED with learning gain81.93 / 39.3081.26 / 38.84
Gain over no-learning-gain control+2.22 / +1.54+1.55 / +1.08

Gain over no-learning-gain control gives +1.55 / +1.08 SS/CQ points. AgriSEED remains ahead of both controls in this experiment. The gain over the frozen solver is +4.65 / +3.79 points, retaining 87% of the SS gain and 89% of the CQ gain.

External evaluation on full AgroBench

The external evaluation uses the full AgroBench benchmark. We follow the five-option exact-match protocol and macro-average accuracy across all seven tasks (Shinoda et al., ICCV 2025).

Full AgroBench macro-average accuracy

ModelMacro accuracy (%)Value type
QwenVLM-7B64.15Published full-set result
QwenVLM-72B70.66Published full-set result
AgriSEED (Qwen2.5-VL-7B)71.3Our result
Gemini 1.5 Pro72.24Published full-set result
GPT-4o73.45Published full-set result

Published scores are from Shinoda et al. (ICCV 2025, Table 2, Overall (all)). The AgriSEED score falls just above QwenVLM-72B and below Gemini 1.5 Pro and GPT-4o.

Thank you for your clarification and additional experiments. However, I remain somewhat unconvinced, as the observed learning gain appears relatively limited. Judging from the comparison across iter 1, iter 2, and iter 3, it seems that further iterations could still be pursued, leaving my initial concerns not yet fully addressed.

Reviewer_ruvEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes AGRISEED, a learning-to-teach framework for agricultural vision–language models. A questioner generates image-grounded question–answer pairs; a frozen visual judge screens them; and a temporary LoRA update to a solver estimates each reliable example’s teaching value from changes in answer likelihood on generated probes. The resulting learning-gain signal, combined with format and answer-quality rewards, trains the questioner using GRPO. The trained questioner then produces filtered supervision for a separate solver-training stage. On AgroCoT, experiments with Qwen2.5-VL-7B, Qwen3-VL-8B, and InternVL2-8B report improvements over base models on semantic similarity and chain-of-thought quality, and comparisons with supervised GRPO and an adapted SPICE baseline. The evaluation uses the full benchmark, including images exposed during adaptation; the authors characterize this as transductive and do not claim it is an image-disjoint test.

Critical strengths
  1. A distinct teaching-value signal for question generation.

    • AGRISEED goes beyond candidate correctness or solver difficulty by measuring whether a temporary update improves performance on probe examples.
    • This is a relevant conceptual contribution to synthetic supervision and self-play.
  2. Broad empirical coverage and informative ablations.

    • The paper evaluates three model backbones across five agricultural task dimensions and reports ablations of the learning-gain reward and probe composition.
    • The reported results are consistent enough to motivate further investigation of the approach.
Critical weaknesses
  1. Limited methodological novelty.

    • AGRISEED combines established ideas, including teacher–student feedback, questioner–solver self-play, GRPO, and visual quality filtering.
    • Its main distinction is using probe-loss change after a temporary LoRA update as the questioner’s reward.
    • The paper does not sufficiently establish that this signal represents a substantial advance over prior learning-to-teach methods or simpler alternatives, such as uncertainty- and success-rate-based feedback.
    • The ablations show that removing the signal hurts performance, but do not establish its novelty or superiority over these alternatives.
    • Overall, the contribution reads primarily as a domain-specific extension of existing methods.
  2. Evaluation and reproducibility concerns.

    • The evaluation includes images used during adaptation, so the results do not establish generalization to unseen images.
    • The sampling description is also inconsistent: the stated 10% protocol does not match some counts in Table 3 and Appendix A.
    • Finally, the method relies heavily on a frozen judge, but provides no independent validation of its agricultural judgments.
Questions for authors
  1. Novelty and positioning: What is the precise methodological contribution that distinguishes AGRISEED from prior learning-to-teach, student-feedback, curriculum-learning, and multimodal self-play approaches? Please clarify which component is novel in principle, rather than specific to the agricultural application or implementation.

  2. Comparison with alternative teaching signals: Did the authors compare the proposed temporary-update gain reward with simpler alternatives, such as solver uncertainty, success-rate-based difficulty, or held-out probe performance without a candidate-specific update? If not, what evidence supports the claim that the proposed signal, rather than questioner training or quality filtering alone, is responsible for the gains?

  3. Sampling protocol: What exact sampling unit, procedure, and adaptation manifest were used? Please reconcile the Section 4.1 description, Table 3 targets, and the alternative totals in Appendix A.

  4. Seen versus unseen images: Can results be reported separately for adaptation-exposed images and images not used during adaptation? If these results are unavailable, please clarify that the reported findings are limited to the transductive setting.

  5. Judge reliability: Was the frozen judge’s screening and scoring compared against independent human assessments or another external correctness measure? If so, please report the audit procedure and agreement results. If not, how do the authors assess the risk that judge errors influence both the teaching signal and the final solver reward?

1 comment

Teaching value. AgriSEED estimates the utility of each generated QA pair from the effect of one temporary solver LoRA update on screened probes, including probes from other images. We compare length-normalized reference-answer log-likelihood on the same probes before and after the update, reset the adapter for each candidate, and use the gain as a term in the questioner's GRPO reward. This uses ideas from learning-to-teach (Fan et al., ICLR 2018) and gradient-based sample reweighting (Ren et al., ICML 2018). The contribution is using each generated pair's effect on answers to other images to train a visual question generator.

Comparison with a simpler reward. We compare temporary-update gain with a reward based on solver success rate, a difficulty signal related to those used in multimodal self-play (He et al., VisPlay, CVPR 2026). The comparison uses Qwen2.5-VL-7B and the same frozen judge, image pool, other reward terms, and training budgets. Both arms use the same Phase 2 difficulty filter and the same held-out AgroCoT images.

Reward comparison on held-out AgroCoT images

Phase 1 conditionSSCQSS gain vs BaseCQ gain vs Base
Frozen base76.535.0——
Original reward80.738.2+4.2+3.2
Learning gain81.939.2+5.4+4.2

The gains over the frozen base are 29% larger for SS and 31% larger for CQ. Each percentage uses the original reward's gain over the base as its denominator.

Sampling and held-out evaluation. The 10% and 20% sampling targets came from different experimental versions. All results in the paper use the 10% target. We will correct the sampling entries in Table 3 and Appendix A in the revision to reflect that version. We use the 10%-target split for adaptation and reserve the remaining 90% of AgroCoT for held-out testing. We group all questions and augmented views by original image and exclude the held-out images from all adaptation stages, including probe feedback. AgriSEED takes only images from the AgroCoT adaptation split. It generates the training QA pairs and uses feedback from a frozen visual judge. AgriSEED's held-out scores are close to the full-set scores reported in the paper, supporting generalization to unseen images within AgroCoT. For Qwen2.5-VL-7B at Iteration 3, the SS/CQ values are 82.2/39.5 on the exposed subset and 81.9/39.2 on the held-out subset.

Human judge agreement. We randomly sampled 10% of the audit pool, which consisted of QA pairs generated during training. Human reviewers and the frozen Qwen3.5-27B judge agreed on 92% of the reviewed pairs.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.