This study investigates the extent to which multimodal large language models (MLLMs) demonstrate physical reasoning capabilities in dynamic, visually grounded environments. We evaluate whether using images as context in a prompt can enhance the ability of MLLMs to simulate and predict the outcomes of physical interactions. Using Science Birds, an Angry Birds-like physicsbased platform, we design a suite of tasks that probe core competencies in binary, comparative stability, and forward simulation using visual and textual inputs. Our evaluation shows that MLLMs can perform visual reasoning tasks with quantifiable accuracy, especially when making predictions based on image-rich input. However, their performance varies considerably depending on the specific model architecture and the type of input modality used. These findings highlight that incorporating both visual and textual data is crucial for accurate physical inference. Nonetheless, the evaluated MLLMs still have substantial limitations in structured visual reasoning tasks. By systematically analyzing the strengths and weaknesses of different models, our work provides practical guidance for advancing MLLM-based physical reasoning and supports the development of future benchmarks and competitions in this area, including contributions to the LLMs4PCG challenge. We make our source code and raw data available for future research.11https://anonymous.4open.science/r/cog2025-game-physics-eval/.

This paper presents a multi-year evaluation (2022-2024) of winning entries from the DareFightingICE Sound Design Competition, analyzing the interplay between audio aesthetics and gameplay functionality for visually impaired players and the Blind AI agent. Through user studies ($\mathrm{n} {=} {2 6}$and$\mathrm{n} {=} {2 1}$) and a comprehensive ablation study of the 2024 winning sound design, we reveal three key findings: (1) While winning entries consistently achieved high aesthetic ratings, scores plateaued despite year-on-year improvements in Blind AI performance, indicating a decoupling of these dimensions; (2) Systematic muting identified “critical” sounds whose absence severely impaired AI decision-making, whereas static background music introduced detrimental noise; (3) Drastically reducing sound design to${1 0 \%}$of sound effects degraded both AI win ratios and human aesthetic perception. These results demonstrate that practical accessible audio requires prioritizing unambiguous cues for frequent actions, minimizing sonic redundancy like non-adaptive BGM, and validating sound designs through dual AI/human evaluation. We provide actionable strategies for designers targeting future competitions and inclusive gaming.

This paper introduces Sonic Doom, an accessibility focused enhancement of the ViZDoom First-person shooter (FPS) platform, integrating an advanced sound design and two aim-assist systems: Auto Aim (automated crosshair adjustment) and Sonic Aim (audio feedback for target proximity). To tackle the issue of FPS games remaining largely inaccessible to visually impaired players (VIPs) due to the game's reliance on visual cues for navigation and combat. We evaluate our approach through a dual-method framework: (1) AI agents trained to play blindly using only an audio input, and (2) human participants in both sighted and blindfolded conditions. Results demonstrate that enhanced sound design improves navigation efficiency (e.g., faster maze completion by the AI agent) and combat accuracy (e.g., higher enemy kill rates in human trials). While Auto Aim achieved superior objective performance, subjective evaluation revealed a strong user preference for Sonic Aim, emphasizing the need to balance assistance with player autonomy. Our AI-driven evaluation framework, the first of its kind in FPS accessibility research, provides scalable, objective metrics for assessing sound designs. This work advances accessible game design by empirically validating sonification techniques in combat-intensive scenarios and establishing a methodology for future research in multi-modal game accessibility.

This paper presents our work to enhance the background music (BGM) in DareFightingICE by adding adaptive features. The adaptive BGM consists of three different categories of instruments playing the BGM of the winner sound design from the 2022 DareFightingICE Competition. The BGM adapts by changing the volume of each category of instruments. Each category is connected to a different element of the game. We then run experiments to evaluate the adaptive BGM by using a deep reinforcement learning AI agent that only uses audio as input (Blind DL AI). The results show that the performance of the Blind DL AI improves while playing with the adaptive BGM as compared to playing without the adaptive BGM.

This demo paper presents a deep reinforcement learning blind agent (blind agent) that uses only sound as the input on the DareFightingICE platform at the DareFightingICE Competition in IEEE CoG 2022 and 2023. A past study used reinforcement learning to train a blind agent and used it as the sample for the AI track of DareFightingICE Competition. However, this approach suffered from significant time expenses due to training the blind agent to recognize sounds and act in an end-to-end manner. To address this challenge, we propose a novel approach for blind agents where sound recognition and decision-making tasks are trained separately.

This paper proposes a delay mechanism to mitigate the impact of latency differences in the gRPC framework—a high-performance, open-source universal remote procedure call (RPC) framework—between different programming languages on the performance of agents in DareFightingICE, a fighting-game research platform. The study finds that gRPC latency differences between Java and Python can significantly impact real-time decision-making. Without a delay mechanism, Java-based agents outperform Python-based ones due to lower gRPC latency on the Java platform. However, with the proposed delay mechanism, both Java-based and Python-based agents exhibit similar performance, leading to a fair comparison between agents developed using different programming languages. Thus, this work underscores the crucial importance of considering gRPC latency when developing and evaluating agents in DareFighting-ICE, and the insights gained could potentially extend to other gRPC-based applications.

This paper presents a new competition-at the 2022 IEEE Conference on Games (CoG)- called DareFightingICE Competition. The competition has two tracks: a sound design track and an AI track. The game platform for this competition is also called DareFightingICE, a fighting game platform. DareFightingICE is a sound-design-enhanced version of FightingICE, used earlier in a competition at CoG until 2021 to promote artificial intelligence (AI) research in fighting games. In the sound design track, participants compete for the best sound design, given the default sound design of DareFightingICE as a sample, where we define a sound design as a set of sound effects combined with the source code that implements their timing-control algorithm. Participants of the AI track are asked to develop their AI algorithm that controls a character given only sound as the input (blind AI) to fight against their opponent; a sample deep-learning blind AI will be provided by us. Our means to maximize the synergy between the two tracks are also described. This competition serves to come up with effective sound designs for visually impaired players, a group in the gaming community which has been mostly ignored. To the best of our knowledge, DareFightingICE Competition is the first of its kind within and outside of CoG.

This paper presents a deep reinforcement learning agent (AI) that uses sound as the input on the DareFightingICE platform at the DareFightingICE Competition in IEEE CoG 2022. In this work, an AI that only uses sound as the input is called blind AI. While state-of-the-art AIs rely mostly on visual or structured observations provided by their environments, learning to play games from only sound is still new and thus challenging. We propose different approaches to process audio data and use the Proximal Policy optimization algorithm for our blind AI. We also propose to use our blind AI in evaluation of sound designs submitted to the competition and define two metrics for this task. The experimental results show the effectiveness of not only our blind AI but also the proposed two metrics.