Mova · papers by stage
world → answer
Answering from the world, with a reason: school tests, reading questions, word problems. The 11 works of Mova's own list that this stage uses; measured marks a work whose effect Mova measured in its own experiments (6 here). Each title opens our card: the summary, how Mova uses the work, its references and who cites it. What the stage does is on Mova's pipeline page; every stage is on the main list.
- Towards AI-Complete Question Answering: A Set of Prerequisite Toy TasksmeasuredJason Weston et al. · 2015
What Mova took:
world/src/babi.rs— theworld babicommand solves all 20 tasks with a reader built on theenparse tree, verb classes, and a first-levelrelationblock, without training on the questions. - Evaluating Models' Local Decision Boundaries via Contrast SetsmeasuredMatt Gardner et al. · 2020
What Mova took:
world/src/babi.rs(CONTRASTtable andcontrast(); commandworld babi-contrast): the bAbI story reader is tested on contrast variants of the bAbI tasks. These keep the same stories but swap in new names, places and… - SPARTQA: A Textual Question Answering Benchmark for Spatial ReasoningmeasuredRoshanak Mirzaee et al. · 2021
What Mova took:
global/seeds/shortcuts/scene.md— a first-level knowledge seed for scene descriptions: categories of colours, sizes and shapes, generic nouns ("thing", "object", "one"), synonyms of above/below, and the relations near, far and… - StepGame: A New Benchmark for Robust Multi-Hop Spatial Reasoning in TextsmeasuredZhengxiang Shi, Qiang Zhang, Aldo Lipani · 2022
What Mova took:
world/src/stepgame.rs— theworld stepgamecommand evaluates Mova's world model on this benchmark. - STaR: Bootstrapping Reasoning With ReasoningmeasuredEric Zelikman, Yuhuai Wu, Jesse Mu, Noah D. Goodman · 2022
What Mova took:
math/src/bin/mathsolve.rs, flag--star N: GSM8K training problems without an exact gold tree go into a pool; in each round the current (averaged perceptron) solver proposes its top-5 expression trees, and a tree whose value… - How Is LLM Reasoning Distracted by Irrelevant Context? An Analysis Using a Controlled BenchmarkmeasuredMinglai Yang et al. · 2025
What Mova took:
math/src/bin/mathsolve.rs, flag--augment: following the idea of training on noisy problems (together with Anantheswaran et al. 2024), each training problem gets a copy with a sentence containing a number from another problem… - A Corpus for Reasoning About Natural Language Grounded in Photographs (NLVR2)Alane Suhr et al. · 2019
What Mova took:
global/seeds/shortcuts/scene.md— the first-level "scene" seed… - BLiMP: The Benchmark of Linguistic Minimal Pairs for EnglishAlex Warstadt et al. · 2020
What Mova took:
en/seeds/errors/attraction-of-phrase.md— the rulesen.errors.attraction-ofanden.errors.attraction-of-auxcopflag agreement attraction through an of-phrase (A pattern of arrests indicate…), excluding quantity nouns… - Are NLP Models really able to Solve Simple Math Word Problems?Arkil Patel, Satwik Bhattamishra, Navin Goyal · 2021
What Mova took: SVAMP is the main benchmark of the math solver in
math(the "steps" model and its judge); nearly every idea tested in the solver is reported as an SVAMP delta. - Fantastic Questions and Where to Find Them: FairytaleQA – An Authentic Dataset for Narrative ComprehensionYing Xu et al. · 2022
What Mova took:
world/src/ftqa.rs: loader for the FairytaleQA CSV (copy from Hugging Face, Apache-2.0 per its card); it reconstructs full stories from per-question sections and splits sentences. Used only for evaluation, never for training. - Evaluating large language models for the tasks of PoS tagging within the Universal Dependency frameworkMateus Machado, Evandro Ruiz · 2024
What Mova took:
en/seeds/errors/llm-annotator-failure-modes.md— failure mode 3 ("labels outside the tagset"): Maritaca produced 93 distinct labels instead of 17 UPOS tags. This supports Mova's practice of passing every LLM-drafted annotation…