First A submission (AAMAS): how much theory is enough when your experiments went sideways? [D]
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一位博士生正在为其首次 AAMAS 投稿寻求建议，因为实验结果仅部分支持其初始假设。他们面临理论构建的挑战，由于在看到结果后才构建理论，陷入了“HARKing”的困境。问题包括理论与特定训练输出而非结构属性相关，以及代码库中发现错误的隐藏参数。他们询问 A* 级别会议对经验性 MARL 论文的理论要求，以及如何在出版压力下从“HARKing”中恢复。
Hi everyone,
2nd-year PhD candidate here staring down my first A* submission deadline (AAMAS 2027). I could really use some perspective on theory expectations, especially since I think I’ve methodologically painted myself into a corner.
The setup
My project started with a clean hypothesis: if architecture X is more robust than Y to perturbation A, and B is a strictly harder version of A, then the X > Y ordering should hold under B as well. I isolated three variables I suspected were driving the effect, ran experiments, and… got results that only partially support the hypothesis, with clear boundary conditions.
Where I got stuck
Trying to explain the “why” mathematically sent me down a theory rabbit hole. I ended up with two bad options:
- Claims tied to specific training outputs rather than structural/architectural properties, or
- Weak, hand-wavy speculations that feel like post-hoc rationalizations.
I’m pretty sure I fell into HARKing.. I started building theory after seeing the results instead of deriving predictions beforehand.
Furthermore, my codebase is built on an undocumented public repo, and I recently found a bunch of hidden parameters set to wrong values for my setting. I’m currently re-running everything, which is why I’m being vague about specifics. My “insights” from the first round are probably garbage.
My actual questions
- For those who’ve reviewed for or published at AAMAS (or similar A* venues): how much formal theory is actually expected for an empirical MARL paper? Is “here’s the phenomenon, here’s the controlled experiments, here’s a plausible but incomplete theoretical sketch” a death sentence?
- If the theory ends up being training-dependent rather than structural, is that a sign I should pivot to a lower-tier venue, or can strong empirical characterization + limited theory still fly at A*?
- How do you recover from HARKing mid-project when you’re under pressure to publish in year 3/4 of a 4-year contract?
Any advice on how to salvage the timeline or reframe the narrative would be hugely appreciated.