acceptodds
Withdrawn from ICLR 2027

Structure-optional learning for proactive assistance

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Proactive assistants advance shared tasks without waiting for explicit requests. Their actions must respect task prerequisites and allow parallel progress alongside a person. Task graphs encode these prerequisites and parallel branches but may be unavailable at inference. We introduce SOLPA (Structure-Optional Learning for Proactive Assistance), which trains a Planner using task-graph supervision. Completion-conditioned questions assume an unfinished step is complete while keeping the person’s current step fixed. Graph-derived answers supervise its next assistance action conditioned on completed-step history. Randomly omitting graph text during training lets the same parameters support inference with and without graphs. On ProAct-75 with predicted task state and annotated completed-step history, our 8B system achieves 33.37% step-success (SS) and 24.10% parallel-action rate (PA), improving over same-backbone Direct by 10.01 and 2.49 percentage points, respectively. Under the same scorer, the newly evaluated strong closed-source models reach at most 14.33% PA without task-specific training. With parameters fixed, the Planner retains 23.94% PA without graph text or its action-name constraint. Our code is available at https://github.com/only4anonymous/SOLPA.git.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language Models, Task Planning, Proactive Assistance

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_CVBoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors introduced SOLPA, a method for training video-based proactive assistants that can select parallel helpful actions without waiting for explicit human requests. The authors used task-graph supervision during training while making graph input optional at inference time. The method introduces two training techniques: completion-conditioned QA, which creates counterfactual supervision by hypothetically marking steps as complete to encourage action reselection, and graph-context dropout, which randomly omits graph text during training to reduce inference-time dependence on structured input like graphs. The system tracks perception (identifying task state from video) and planning (selecting assistance actions) separately, and is trained via LoRA-based supervised finetuning on the ProAct-75 benchmark. The 8B configuration achieves 33.37% step-success and 24.10% parallel-action rate, outperforming the same-backbone Direct baseline by 10 and 2.5 percentage points, respectively.

Critical strengths
  1. The task is well established and of good interest.
  2. Whatever experiments were done were done in a mostly sound way.
  3. The completion-conditioned QA and graph-context dropout insights are interesting.
Critical weaknesses
  1. The writing could be significantly improved, especially the introduction and motivation sections. The current introduction jumps into explaining what the current methods do and what the authors are proposing. However, the exact motivation and research questions could have been explained more explicitly and clearly. Currently, the introduction reads more like: we have a task and a dataset and here's what we do to improve the performance.

  2. One of the paper's main pitch is removing graph dependency at inference, but the key advantage of inference-time graphs was domain adaptability without retraining. But I didn't find any analysis on the generalizability of the proposed approach. No cross-domain or structurally diverse graph experiments are done.

  3. The training approach seems pretty much SFT-only; no RL-based mechanism was explored for the model to learn from planning mistakes. For a task requiring structural reasoning about parallel workflows, SFT likely teaches pattern matching over labels rather than genuine task-graph reasoning.

  4. The authors studied only a single dataset (ProAct-75). No evidence of transfer to other proactive assistance settings, real-world deployment, or benchmarks with different task structures. This significantly limits the scope of the method and contribution.

  5. The analysis is insufficient across model scales. The 8B model gets the headline results, but there's thin analysis of why different backbones (3B, 7B, 8B) respond differently to the training. Base model capabilities before finetuning are under-explored as they could have been explored with more test-time approaches.

  6. [Minor] The related work section on structured planning is underexplored. The authors cited mostly earlier work but more recent studies on constrained generation and structured planning with LLMs are missing. A more thorough survey of this space would better contextualize the proposed approach.

Questions for authors

Please refer to the weaknesses section.

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

This paper presents a new training framework, called SOLPA, that trains better proactive agents, i.e., agents that can proactively do tasks for humans. It seems to be built heavily upon a prior work called "ProAct-Helper", and seem to have done some improvement on it. It conducts experiment on ProAct-75

Critical strengths

None.

I cannot understand a big chunk of the methodology (see below), and as such, I cannot judge its novelty. Experimental results also seem to suggest very limited improvement (see below) as presented.

Critical weaknesses

I find the methodology very hard to understand. I have attempted to read it many times, but it seems too dense without clearing defining the task and terms. For instance, here:

We use ProAct-Helper’s graph-based action preference offline to recompute the recommended parallel action after marking a candidate step complete, while keeping the person’s current step fixed (Figure 3). We pair this supervision with graph-context dropout, which omits graph inputs from a fraction of planning examples without changing their targets. The same Planner supports graph-on inference with graph text and action names restricted to the graph’s vocabulary, and graph-off inference without either.

Who generates a candidate step? What exactly is a task graph? What is graph inputs? What is "graph-on inference"? What is "graph text"? What is "graph's vocabulary"? It might be possible if someone is familiar with ProAct-Helper, which seems to be a closely related work? In either case, this needs to be better defined and as written and presented, I cannot understand what exactly is the procedure and hence cannot accurately assess what is the novelty.

The experimental result also seem not impressive, or perhaps it is not presented well. Looking at Table 1 for instance, I don't understand by looking at the Acc or F1 columns, why SOLPA (Head) is better than ProAct-Helper. It seems to be just 1-2 points better based on Q-3B or Q-7B.

Questions for authors

I'd consider improving the score if you can present a paper that is easier for a human to understand. At least try to explain the methodology and your novelty better.

Reviewer_QK33OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces SOLPA, a framework for proactive assistance that uses task-graph supervision during training while allowing the graph to be omitted at inference. The main idea is to teach a planner to select useful parallel actions without relying on explicit graph-based search at test time.

Critical strengths
  1. A practical formulation of structure-optional planning is interesting. Learning from task graphs without requiring them at inference is an interesting direction for proactive assistance, where manually specified procedural structures may not always be available. Table 2 provides a useful controlled comparison.

  2. Table 3 separates the contributions of history, graph context, and the QA objective. In particular, completion-conditioned QA improves graph-on/off PA by 3.84/4.94 points over the history-and-graph baseline, with reported confidence intervals excluding zero. The matched comparison with current-state QA provides further evidence that reasoning about hypothetical progress is useful.

Critical weaknesses
  1. The reliance on annotated history seems to create a limiting central deployment claim. The main system-level results use predicted task state but annotated completed-step history. This is an important form of privileged information, particularly because tracking completed steps across task branches is itself a difficult part of proactive assistance. Although the paper studies oracle versus predicted perception, it does not (as per my understanding) establish how the complete system performs when history must also be inferred online. Consequently, graph-free inference is demonstrated under a considerably more favorable setting than a fully autonomous assistant would encounter. I feel this leads to a biased result.

  2. The empirical advantage over existing approaches is not fully established and the paper seems to imrove over a directly same backbone without generalization tests.

Questions for authors

a. How much does performance degrade when completed-step history is obtained from model predictions rather than annotations?

b. Does graph-off performance hold for unseen task graphs, altered prerequisites, or novel step combinations? The current results show that graph text can be omitted, but it is less clear whether the model has learned transferable procedural reasoning rather than task-specific action preferences.

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

Related papers

Discussion (0)

Sign in to comment.