CheeseBench: Evaluating Large Language Models on Rodent Behavioral Neuroscience Paradigms

Zacharie Bugaud

cs.AIApr 12, 2026

CheeseBench: Evaluating Large Language Models on Rodent Behavioral Neuroscience Paradigms

Authors: Zacharie Bugaud

Organizations: Astera Institute

Abstract

We introduce CheeseBench, a benchmark that evaluates large language models (LLMs) on nine classical behavioral neuroscience paradigms (Morris water maze, Barnes maze, T-maze, radial arm maze, star maze, operant chamber, shuttle box, conditioned place preference, and delayed non-match to sample), spanning six cognitive dimensions. Each task is grounded in peer-reviewed rodent protocols with approximate animal baselines. The agent receives a unified system prompt with no task-specific instructions and must discover goals purely from ASCII text observations and reward signals, much like a rodent placed into an unfamiliar apparatus. We evaluate six open-weight LLMs (3B to 72B parameters) on text-based ASCII renderings and compare against both a random baseline and a graph-based reinforcement learning agent. Our best model (Qwen2.5-VL-7B) reaches 52.6% average success on ASCII input, compared to 32.1% for random agents and 78.9% for approximate rodent baselines. We find that (1) scaling beyond 7B yields diminishing returns, (2) longer context history degrades performance, (3) chain-of-thought prompting hurts rather than helps, and (4) a vision-language architecture provides an advantage at 7B but hurts at 32B. Because the same model's performance ranges from 20% to 57% depending on interface parameters alone, these results characterize the agent-plus-interface system, not the model in isolation. Under this unified zero-shot ASCII protocol, current open-weight LLM agents remain well below approximate rodent reference values, particularly on tasks requiring spatial navigation and within-trial state tracking.

Figures & tables

Figure 1: ASCII renderings of all nine CheeseBench environments as seen by the model. Walls are shown as # , traversable areas as ˜ or . , goals as G , and the agent arrow indicates position and facing direction.

Environment	Cognitive Dimension	Species	Trials	Steps
MorrisWaterMaze	Spatial Learning	C57BL/6	20	500
BarnesMaze	Spatial Learning	C57BL/6	16	300
StarMaze	Spatial Learning	C57BL/6	40	300
TMaze	Ego. Navigation	C57BL/6	40	200
RadialArmMaze	Working Memory	C57BL/6	20	400
DNMSTask	Working Memory	Long-Evans	100	50

Table 1: CheeseBench environments. Animal baselines are from published rodent studies. Trials = evaluation trials; Steps = max steps/trial.

Figure 2: The three ASCII observation modes for the Star Maze. ASCII_2D provides a full top-down text map; ASCII_2D_FPV shows an egocentric cropped view centered on the agent; ASCII_3D renders a pseudo-3D first-person perspective with depth shading.

Figure 3: Per-environment success rates (ASCII text input, best of three view modes per environment) for six open-weight LLMs, Tabular-QL, and random baseline compared to approximate rodent reference values (diamonds). The Radial Arm Maze remains essentially unsolved by all open-weight agents; the Frontier Models subsection shows that frontier models close this gap.

Figure 4: Cognitive profile across six dimensions. The LLM and Tabular-QL agent show complementary strengths: the LLM handles a broader range of tasks while Tabular-QL excels at conditioning through direct reinforcement learning.

Figure 5: Scaling behavior: performance saturates beyond 7B under this protocol.

Figure 6: Ablation studies (all on ASCII text input) using Qwen2.5-VL-7B (32B for vision ablation). (a) Shorter history is better. (b) Default prompt beats CoT. (c)

k{=}4

is the sweet spot. (d) Vision encoder helps at 7B but hurts at 32B.

Model (tier)	Overall
Claude Opus 4.7 (frontier)	86.7
GPT-5.2-Codex (frontier)	82.2
Claude Opus 4.6 (frontier)	82.2
Claude Sonnet 4.6 (frontier)	81.1
GPT-5.2 (frontier)	75.6
Claude Haiku 4.5 (frontier)	70.0

Table 2: Frontier-model overall success rate (%) compared to the best open-weight model and the rodent reference. “Overall” is the mean of per-environment best-view-mode success rate. Trial counts: frontier models 270 per run (GPT-4.1: 81); open-weight 1352 per run.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Environment	P / S / C
MWM	P: Hidden goal, distal landmarks, multiple starts. S: Continuous swim $\to$ grid; no currents. C: Allocentric spatial learning.
Barnes	P: Escape hole among decoys, aversive motivation. S: Light/noise $\to$ step penalty; 3D $\to$ 2D. C: Spatial reference memory.
Star	P: Radial corridors, strategy choice. S: Hidden goal $\to$ visible marker; continuous motion $\to$ grid; corridor width. C: Spatial strategy selection.
T-Maze	P: T-corridor, forced choice, reward at one arm. S: Olfactory/tactile cues; continuous turning. C: Egocentric navigation.
RAM	P: 8 arms (4 baited), revisit tracking. S: Odor trails; arm length variation; no explicit revisit penalty. C: Working memory across choices.
DNMS	P: Sample–delay–choice, non-match rule. S: Object features $\to$ symbols; step-based delay. C: Working memory + rule learning.

Appendix

Table 4: Environment fidelity summary. P : preserved from rodent protocol. S : simplified away. C : cognitive demand retained.

Env	Approx.	Sess.	PMC	Notes
MWM	85%	5	PMC3259155	Platform acquisition, C57BL/6 mice, 5 trials/sess. Metric is path length; mapped to binary (reach platform within step limit).
Barnes	80%	5	PMC1783636	12 holes, 4 trials/sess. Metric: primary errors $\to$ binary correct-hole.
Star	80%	10	PMC3695082	5-arm maze, 4 trials/sess. Metric: reach goal arm via correct path.
T-Maze	80%	4	PMC3399492	Forced alternation, 10 trials/sess. Metric: correct arm on free choice.
RAM	70%	6	PMC4030456	8 arms, 4 baited. Collect all rewards with 0 WM errors. Conservative; original rats scored higher on simpler metrics.
DNMS	80%	30	PMC3982138	Touchscreen TUNL task (rats). Criterion is 80% correct on large separations at short delay.

Appendix

Table 5: Rodent baseline construction. Approx. = approximate asymptotic success rate used in figures. Sessions = number of training sessions modeled in the learning curve. PMC = PubMed Central source article for protocol.

Explore similar work

Feb 20, 2026

Rodent-Bench

We present Rodent-Bench, a novel benchmark designed to evaluate the ability of Multimodal Large Language Models (MLLMs) to annotate rodent behaviour footage. We evaluate state-of-the-art MLLMs, including Gemini-2.5-Pro, Gemini-2.5-Flash and Qwen-VL-Max, using this benchmark and find that none of these models perform strongly enough to be used as an assistant for this task. Our benchmark encompasses diverse datasets spanning multiple behavioral paradigms including social interactions, grooming, scratching, and freezing behaviors, with videos ranging from 10 minutes to 35 minutes in length. We provide two benchmark versions to accommodate varying model capabilities and establish standardized evaluation metrics including second-wise accuracy, macro F1, mean average precision, mutual information, and Matthew's correlation coefficient. While some models show modest performance on certain datasets (notably grooming detection), overall results reveal significant challenges in temporal segmentation, handling extended video sequences, and distinguishing subtle behavioral states. Our analysis identifies key limitations in current MLLMs for scientific video annotation and provides insights for future model development. Rodent-Bench serves as a foundation for tracking progress toward reliable automated behavioral annotation in neuroscience research.

Thomas Heap, Laurence Aitchison, Emma Cahill +1

University of Bristol, Bristol, UK

Animal Behavior Analysis Video-Language Model Evaluation Multimodal Large Language Models Multimodal Model Evaluation

Ask PDF

Apr 9, 2026

ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models

Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval. This gap is critical: effective assistants must automatically apply learned procedures or avoid failed actions without explicit reminders. We introduce ImplicitMemBench, the first systematic benchmark evaluating implicit memory through three cognitively grounded constructs drawn from standard cognitive-science accounts of non-declarative memory: Procedural Memory (one-shot skill acquisition after interference), Priming (theme-driven bias via paired experimental/control instances), and Classical Conditioning (Conditioned Stimulus--Unconditioned Stimulus (CS--US) associations shaping first decisions). Our 300-item suite employs a unified Learning/Priming-Interfere-Test protocol with first-attempt scoring. Evaluation of 17 models reveals severe limitations: no model exceeds 66% overall, with top performers DeepSeek-R1 (65.3%), Qwen3-32B (64.1%), and GPT-5 (63.0%) far below human baselines. Analysis uncovers dramatic asymmetries (inhibition 17.6% vs. preference 75.0%) and universal bottlenecks requiring architectural innovations beyond parameter scaling. ImplicitMemBench reframes evaluation from "what agents recall" to "what they automatically enact".

Chonghan Qin, Xiachong Feng, Weitao Ma +2

The University of Hong Kong · Harbin Institute of Technology

Memorization in Language Models Memory Agent Benchmarks

Ask PDF

Mar 16, 2026

BrainBench: Exposing the Commonsense Reasoning Gap in Large Language Models

Large language models (LLMs) achieve impressive scores on standard benchmarks yet routinely fail questions that any human would answer correctly in seconds. We introduce BrainBench, a benchmark of 100 brainteaser questions spanning 20 carefully designed categories, each targeting a specific commonsense reasoning failure mode in LLMs. Categories range from implicit physical constraints ("Should I walk or drive my rental car to the return lot?") to semantic scope tricks and default assumption hijacks. We evaluate eight frontier models -- four from the Claude family and four from the GPT family -- using a zero-shot protocol with 10 independent runs per question. The best model, Claude Opus 4.6 with extended thinking, achieves only 80.3% accuracy; the worst, GPT-4o, scores 39.7%. Even top-performing models exhibit a 6-16 percentage-point gap between accuracy and consistency, revealing stochastic reasoning. Cross-lingual evaluation in Chinese shows most models degrade by 2-8 percentage points, confirming that these failures reflect reasoning deficits rather than language-specific artifacts. BrainBench provides a fine-grained diagnostic tool for identifying where and why LLMs substitute surface heuristics for genuine commonsense reasoning.

Yuzhe Tang

Georgia Institute of Technology

LLM Evaluation Implicit Reasoning in Language Models Reasoning Consistency in Language Models

Ask PDF