{"id":"4d200b1d-02ec-400e-bf8d-57e13698031f","arxiv_id":"2501.00078","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Ray-cast sensor-based imitation learning produces compute-efficient, human-like bots for a tactical shooter, with a 14.9M-parameter model running at 9.59 ms per decision on CPU.","lead":"This paper trains small neural-network game bots for a VALORANT-like 2v2 shooter using ray-cast sensors instead of pixels, so the bots run on ordinary CPUs. The authors report the best bot decides in about 10 milliseconds per step and is near chance in a human-versus-bot video classification test.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Believability claim rests on training-set similarity and an underpowered human study, and Table IV contains a mathematically impossible negative JS divergence.","rationale":"The paper's most concrete and defensible contribution is the CPU inference-time comparison, which is direct and internally consistent. The title-level claim of human-likeness, however, is supported only by distributional comparisons against the training data and a small classification study with no control condition and no statistical analysis. The reader's weakest assumption identifies exactly this gap. My reading adds one concrete red flag: Table IV contains a negative JS divergence, which is mathematically impossible and suggests the quantitative similarity analysis is unreliable. That unreliability strengthens the need for a held-out evaluation but does not, by itself, overturn the compute-efficiency result. A held-out split is the minimal check that would determine whether the reported human-likeness is a real property of the policy or an artifact of training-data overlap. Since the reader already issued a CONDITIONAL verdict, my concerns do not move the verdict; they reinforce it. I agree with the reader's weakest assumption rather than proposing a different one, because the believability evidence is the weakest link in the central claim. The lack of released code and data makes independent verification harder, but the proposed held-out analysis is feasible for the authors and would settle the main worry.","tokens_in":17030,"tokens_out":6530,"duration_ms":70173,"concrete_test":"Hold out 20% of human rounds (at the round or player level, not frame level) before training model D, then generate bot-vs-bot data and recompute the JS divergences of Table IV against the held-out human data. If the divergences change materially or the model ranking shifts, the reported human-likeness is an artifact of evaluating on the training distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VIII-B computes JS divergence between bot-generated data and the same 48.6 hours of human data used for training, with no held-out split; distributional similarity therefore partly measures fit to the training distribution, not human-likeness in novel situations. The human study in Section IX uses 20 participants classifying 32 curated 15-second clips, with no non-human baseline and no significance test; a 54% correct bot-identification rate and a 30% 'human' rate are not interpretable without a chance or scripted-bot comparator. In addition, Table IV's Defence/Speed row reports a JS divergence of -0.249 for Model A, which is impossible since JS divergence is non-negative; this indicates an error in the exact table used to select Model D. These issues do not affect the inference-time comparison, but they leave the paper's human-like claim unsupported. The compute-efficiency numbers are the strongest evidence; the believability conclusion is the load-bearing weak point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a sensor-based perception architecture for training human-like bots in a VALORANT-like 2v2 tactical shooter (Lyra:Ascent). Instead of pixel input, the agents use ray-cast visual sensors, directional audio sensors, distance/direction features, and game-state information. Six LSTM-based behavior-cloning models (A-F) are trained on 48.6 hours of human gameplay and evaluated on inference time, distributional similarity to human data (Jensen-Shannon divergence), spatial similarity via heatmaps (EMD/ASD), and a human perception study. The authors report that model D (14.9M parameters) runs at 9.59 ms per decision on a consumer CPU, significantly faster than a comparable CS:GO pixel-based model (24.10 ms), and is the most human-like according to their metrics.","tokens_in":17186,"tokens_out":4528,"duration_ms":44280,"significance":"If the human-likeness claims survive scrutiny, the paper makes a valuable contribution to practical game AI: it demonstrates that compute-efficient, non-pixel sensors can support behavior-cloned agents in a commercial-standard shooter, with concrete and favorable inference-time comparisons. The sensor design is novel and the engineering effort is substantial. However, the believability evidence currently rests on comparisons against the training data itself and on a small human study without a baseline or significance tests; the quantitative model-selection table also contains an impossible negative JS divergence. The compute-efficiency result is the strongest part of the paper, while the human-likeness conclusion is not yet supported.","major_comments":[{"comment":"The distributional similarity analysis compares bot-generated data to the same 48.6 hours of human gameplay used to train the models, with no held-out split. Consequently, the reported JS divergences partly measure how well the models fit their training distribution rather than how human-like they are in novel situations. The conclusion in Section VIII-B that model D is the best-performing model is therefore not supported as a generalization claim; a held-out human dataset or cross-validation is needed before this metric can support the human-likeness claim.","section":"Section VIII-B, Table IV"},{"comment":"The Defence/Speed row reports a JS divergence of -0.249 for model A, which is impossible because Jensen-Shannon divergence is non-negative by construction (Eq. 1). This indicates an error in the computation or in the table itself. Since this table is used to select model D as the best model, the error undermines the quantitative model comparison unless corrected.","section":"Table IV, Defence/Speed row"},{"comment":"The human perception study uses only 20 participants classifying 32 curated 15-second clips, with no non-human baseline (e.g., a scripted bot or random guessing) and no significance tests or confidence intervals. The 54% correct bot identification rate is close to chance and is not interpretable without such statistical context, so the conclusion that the bots are 'convincing' or 'sophisticated' is not statistically supported.","section":"Section IX, Human perception study"},{"comment":"The spatial similarity evaluation (EMD and ASD) also uses heatmaps derived from the human training data as the reference distribution. As with the JS divergence analysis in Section VIII-B, this measures fit to the training set rather than generalization to unseen human behavior. The claim in Section VIII-C that model D is 'the most robust in replicating human-like moving patterns' is therefore not supported for novel situations.","section":"Section VIII-C, Table V"}],"minor_comments":[{"comment":"There is a typo in the contributions list: 'beliavability' should be 'believability'.","section":"Section I, Contributions"},{"comment":"The text spells 'Kullback-Liebler'; the correct spelling is 'Kullback-Leibler'.","section":"Equation 1"},{"comment":"The phrase 'The data was then further cleaned' should be 'The data were then further cleaned' for subject-verb agreement.","section":"Section V"},{"comment":"The inference-time comparison would be easier to interpret if the authors stated how many runs were averaged for each model and whether the measurements were on a single CPU thread; the reported error bars are standard deviations but the number of trials is not given.","section":"Section VIII-A, Table III"},{"comment":"The heatmap figure caption lists only models A, C, D, and E in addition to the human condition; it is unclear why models B and F are omitted from the visualization.","section":"Figure 8"},{"comment":"The paper does not state whether the 20 participants in the human study overlapped with the 28 players whose gameplay was used for training; if there was any overlap, it should be disclosed as a potential familiarity bias.","section":"Section IX-A"}],"recommendation":"major_revision","confidential_remarks":"The core compute-efficiency result is solid and should be retained as the main contribution. The negative JS divergence in Table IV is a clear correctness error that must be fixed, and the lack of a held-out evaluation is a methodological gap that needs new data or a reanalysis. The human study is underpowered and needs a proper baseline and significance testing; if the authors cannot provide these, the human-likeness claim should be substantially weakened. The paper fits the journal's scope, but the believability evidence currently does not support the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core engineering result is real: a ray-cast sensor-based behavior cloning bot for a tactical shooter that runs in 9.59 ms per decision on a desktop CPU, versus 24.10 ms for a pixel-based CS:GO model. That is a concrete, useful advance for commercial game AI. Second, the believability evidence is considerably weaker than the abstract suggests, and Table IV contains a negative Jensen-Shannon divergence, which is impossible. The paper needs revision, not rejection.\n\nWhat is genuinely new: replacing pixel input with a non-uniform 15x15 ray grid, coupling the sensor angles to a discrete 2D aiming action space, and showing that this cuts inference time dramatically. The architecture is sensible, the comparison to Pearce and Zhu is appropriate, and the measured inference times are the strongest part of the paper. That alone is worth publishing if the evaluation is cleaned up.\n\nThe soft spots are structural. Distributional similarity (JS divergence, heatmap EMD) is computed between bot-generated data and the exact 48.6 hours of human data used for training, with no held-out split. Those numbers partly measure overfitting, not human-likeness in novel situations. The human study uses 20 participants, 32 curated clips, no non-human baseline, and no significance test. A 54% correct bot-identification rate, with 30% of bot clips judged human, is not interpretable without a chance comparator or a scripted-bot control. And the negative JS divergence in Table IV (Defence/Speed, Model A) is mathematically impossible; that error sits in the very table used to select Model D. These are fixable, but they leave the central claim unsupported.\n\nI also note no code or data release, which limits reproducibility, though the inference-time comparison is reproducible in principle.\n\nWho is this for? Game AI practitioners, especially those building deployable bots. The paper is a systems contribution, not a new scientific principle, but it is an honest and useful one. The evaluation gaps are addressable, and the compute-efficiency result is solid.\n\nMy recommendation: send it to peer review. A serious referee should ask for a held-out human evaluation, proper baselines, corrected Table IV, and a proper statistical treatment of the Turing-test data. If those are delivered, the paper would be a solid venue contribution.","headline":"Real compute-efficiency contribution, but the human-likeness evidence is undermined by training-set evaluation and a mathematically impossible negative JS divergence.","tokens_in":17857,"tokens_out":2058,"would_cite":true,"duration_ms":21754,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that imitation learning with a pixel-free ray-cast sensor grid can produce human-like tactical-shooter bots that make decisions in single-digit milliseconds on a CPU, with a video Turing test showing 30% of bot clips…","keywords":["imitation learning","behavior cloning","ray-cast sensors","compute-efficient AI","human-like bots","tactical shooter","LSTM","video game AI"],"falsifier":"Compare the bots' behavior against a held-out corpus of human gameplay that was never used for training, using the same Jensen-Shannon divergence and heatmap-distance metrics; if similarity to held-out humans is substantially worse than similarity to the training humans, the human-likeness claim is inflated. A preregistered live-interaction Turing test, where players face the bots in real time with a larger participant pool, would test whether the 30% human-judgment rate survives outside curated clips.","tokens_in":16848,"feed_emoji":"🎮","tokens_out":7477,"duration_ms":67680,"temperature":0.7,"pith_summary":"This paper tries to show that neural-network bots for a commercial-standard tactical shooter can be both human-like and cheap enough to run on a CPU. It replaces pixel-based vision with a small grid of ray-cast sensors that are denser toward the crosshair, and it couples each sensor direction to a matching aiming action. The bots are trained by behavior cloning on 48.6 hours of human gameplay, using networks with long short-term memory (LSTM) to carry context across time. On a desktop CPU, the best model makes a decision in 9.59 ms on average, and in a video Turing test 30% of bot clips were judged human. If these results hold, human-like AI opponents and teammates could be deployed in commercial games without GPU inference or expensive pixel processing.","feed_headline":"Ray-cast sensors put human-like shooter bots on CPU in 9.59 ms","feed_subtitle":"Imitation-learned VALORANT-style bots need no pixels, and a third of judges mistook their clips for human players.","key_machinery":"The load-bearing mechanism is the coupled sensor-action ray grid: 225 range finders arranged non-uniformly, dense near the crosshair and sparse at the periphery, plus one extra ray per important tracked object inside a 90-degree field of view. Each sensor direction has a matching discrete aiming action, so the bot has a direct action to aim at whatever it senses, and the network learns the sensor-to-action mapping without processing pixels. Audio sensors, distance-and-direction vectors, and game-state features are concatenated into the same model, and a long short-term memory (LSTM) layer provides the temporal memory needed for imitation learning from human trajectories via backpropagation through time.","core_discovery":"The central claim is that imitation learning with pixel-free sensors is sufficient to produce believable tactical-shooter bots that meet production latency budgets. Concretely, a 14.9-million-parameter network with convolutional and LSTM layers, fed by a 15-by-15 ray-cast grid over ten object types together with audio and game-state features, reproduces human behavioral distributions for round duration, shots fired, kills, plant and defuse attempts, and spatial heatmaps of map coverage. In the paper's video Turing test, 30% of bot clips were judged human and 16% drew uncertainty, while human clips were correctly identified 60% of the time. The authors conclude that this model best balances computational efficiency, quantitative similarity to human play, and perceived believability.","pith_inferences":["The paper's evaluation likely overstates human-likeness, since the distributional baselines come from the same data the bots were trained on; a held-out human dataset would give a fairer, probably lower, similarity score.","The 15-second curated clips remove the interactive loop; in a live match, players could exploit telltale patterns (such as the reported higher average movement speed) that static clips may hide.","The ray-grid perception scheme is a general design pattern: any 3D game with a spatial state could replace pixel encoders with structured range sensors, trading visual detail for a large cut in inference cost, at the price of losing non-spatial visual information like UI cues and subtle animations.","The method's failure modes, including weak handling of unseen situations, could be addressed with the paper's own suggested extensions such as GAIL or Decision Transformers, and the sensor-action coupling would remain usable in those frameworks."],"forward_implications":["A single CPU thread can run the best model's decisions in 9.59 ms on average, keeping the GPU free for rendering, which matches the resource budget of commercial shooters.","Behavior cloning from human trajectories alone—without reinforcement learning—reproduces human distributions on rounds, shots, kills, and bomb events, suggesting imitation can yield believable agents.","The best model sits near the middle of the size range tested; larger networks (roughly 25M to 36M parameters) do not consistently improve human-likeness and run 2–3 times slower.","The sensor-action coupling idea may transfer to other 3D games where spatial perception, rather than pixel detail, drives decisions.","Bot clips were judged human 30% of the time with 16% uncertainty, a similar ambiguity range to human clips (60% correct, 25% mistakenly called bots)."],"supporting_citations":[{"why":"The CS:GO behavioral-cloning baseline supplies the discrete sensor-aligned action-space idea, the 16-observation-per-second preprocessing rate, and the 24.10 ms pixel-input inference time that the paper's models beat.","marker":"[3]"},{"why":"The long short-term memory unit is the memory mechanism the network architectures use to imitate temporal human trajectories.","marker":"[10]"},{"why":"The imitation-learning survey defines the problem setup and notation the paper's behavior-cloning formulation follows.","marker":"[6]"},{"why":"The study of how people evaluate human-like AI in games motivates the video Turing-test protocol used for believability.","marker":"[22]"},{"why":"The video-game Turing-test variant informs the design of the 15-second clip classification experiment.","marker":"[28]"},{"why":"The backpropagation-through-time method is the training algorithm used with the LSTM layers.","marker":"[29]"}],"fun_headline_variants":["Imitation learning with ray-cast sensors fools a third of judges","Pixel-free shooter bots run on CPU and fool many humans","Human-like bot play from 9ms CPU inference, no GPU needed","Shooter AI that reeks human: 30% Turing test success on CPU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The believability conclusion depends on treating the training set's own statistics as ground truth for 'human-like': the distributional comparisons use the same 48.6 hours of gameplay the models were trained on, with no held-out human data, and the human study relies on 20 participants judging curated 15-second clips rather than interactive play.","fun_headline_variants_meta":{"raw":{"variants":["Imitation learning with ray-cast sensors fools a third of judges","Pixel-free shooter bots run on CPU and fool many humans","Human-like bot play from 9ms CPU inference, no GPU needed","Shooter AI that reeks human: 30% Turing test success on CPU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000811,"raw_usage":{"total_tokens":3561,"prompt_tokens":956,"completion_tokens":2605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2528}},"tokens_in":572,"tokens_out":2605,"duration_ms":19142,"temperature":1.0,"reasoning_tokens":2528,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:07:29.424401+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the bots' behavior against a held-out corpus of human gameplay that was never used for training, using the same Jensen-Shannon divergence and heatmap-distance metrics; if similarity to held-out humans is substantially worse than similarity to the training humans, the human-likeness claim is inflated. A preregistered live-interaction Turing test, where players face the bots in real time with a larger participant pool, would test whether the 30% human-judgment rate survives outside curated clips.","supporting_citations":[{"cited_title":"Counter-strike deathmatch with large-scale behavioural cloning,","cited_arxiv_id":null,"evidence_quote":"The CS:GO behavioral-cloning baseline supplies the discrete sensor-aligned action-space idea, the 16-observation-per-second preprocessing rate, and the 24.10 ms pixel-input inference time that the paper's models beat."},{"cited_title":"Long short-term memory,","cited_arxiv_id":null,"evidence_quote":"The long short-term memory unit is the memory mechanism the network architectures use to imitate temporal human trajectories."},{"cited_title":"Imitation learning: A survey of learning methods,","cited_arxiv_id":null,"evidence_quote":"The imitation-learning survey defines the problem setup and notation the paper's behavior-cloning formulation follows."},{"cited_title":"Navigates like me: Understanding how people evaluate human-like ai in video games,","cited_arxiv_id":null,"evidence_quote":"The study of how people evaluate human-like AI in games motivates the video Turing-test protocol used for believability."},{"cited_title":"How humans perceive human-like behavior in video game navigation,","cited_arxiv_id":null,"evidence_quote":"The video-game Turing-test variant informs the design of the 15-second clip classification experiment."},{"cited_title":"Backpropagation through time: what it does and how to do it,","cited_arxiv_id":null,"evidence_quote":"The backpropagation-through-time method is the training algorithm used with the LSTM layers."}],"review_version":1}