{"id":"d9fc997a-e2b4-498d-a794-62ab64518930","arxiv_id":"2606.06565","paper_version":1,"verdict":"UNVERDICTED","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"AI LOD uses distance to select quantized precision levels of ML motion models for efficient NPC animation in games.","lead":"This paper proposes AI Level of Detail, a framework that switches machine learning models for game character animation between high and low precision versions based on distance from the player camera. A smart generalist might read it to see a practical way to cut compute costs in real-time AI systems while keeping visuals acceptable.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Quantitative mocap metrics do not establish that quantization errors remain imperceptible at assigned distances","rationale":"The reader's weakest assumption directly identifies the missing perceptual link; the abstract's claim of 'negligible perceptible degradation' is unsupported by the described evaluation method. This single gap prevents an unconditional verdict even if the quantitative results are reproducible.","tokens_in":1695,"tokens_out":309,"duration_ms":15639,"concrete_test":"Run a within-subjects perceptual study: render the same motion sequences at the paper's proposed distance bands using each ONNX precision variant, present randomized pairs (FP32 baseline vs. quantized) to 30+ participants under game-like viewing conditions, and measure forced-choice detection rates; if detection for the farthest INT8 band is statistically indistinguishable from chance, the imperceptibility assumption holds.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that FP32/FP16/INT8 tiers can be mapped to distance ranges with degradation that is imperceptible to players. The evaluation on CMU Mocap reports that each tier can be served at its assigned range with negligible perceptible degradation, but this rests on standard numerical motion metrics (e.g., joint-position or velocity error) rather than any perceptual model, rendering simulation, or human-subject study that accounts for distance-dependent visual acuity, screen resolution, or animation artifacts visible only under motion. The distance thresholds themselves appear chosen heuristically to match the observed error tiers without an independent perceptual calibration step.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes 'AI Level of Detail' (AI LOD), a framework that adapts inference precision of a learned human motion model (convolutional sequence-to-sequence from Li et al., exported to ONNX FP32/FP16/INT8 variants) according to the distance between each NPC and the player camera. The central claim is that distance-based selection of these precision tiers can serve as an LOD axis for AI-driven character animation, with evaluation on the CMU Mocap dataset providing evidence that each tier produces negligible perceptible degradation when used at its assigned distance range. The contribution is framed as the conceptual extension of classical geometry LOD to ML inference precision for any context-sensitive AI runtime system.","tokens_in":1840,"tokens_out":488,"duration_ms":17166,"significance":"If the empirical premise holds, the work introduces a practical, low-overhead runtime optimization for AI animation in games by exploiting distance-dependent perceptual tolerance, potentially reducing GPU/CPU load without visible quality loss. It supplies a concrete, immediately implementable example using existing ONNX Runtime quantization and a distance selector. The idea is generalizable beyond motion prediction and receives credit for identifying inference precision as a controllable LOD dimension rather than requiring retraining or architectural changes.","major_comments":[{"comment":"Abstract: the claim that 'evaluation on the CMU Mocap dataset provides initial evidence that each precision tier can be served at its assigned distance range with negligible perceptible degradation' supplies no quantitative error metrics (joint-position, velocity, or acceleration error), no description of how perceptual quality was assessed, and no human-subject or rendering-based validation that accounts for screen resolution, motion parallax, or distance-dependent visual acuity.","section":"Abstract"},{"comment":"Evaluation section (referenced in Abstract): the distance thresholds that map FP32/FP16/INT8 tiers to ranges appear chosen to match observed numerical error tiers without an independent perceptual calibration step, human study, or simulation of in-game viewing conditions, leaving the weakest assumption (imperceptibility at assigned distances) unsupported.","section":"Evaluation"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a concise concept note rather than a full empirical paper; its fit for a graphics journal would improve with the addition of the missing quantitative and perceptual results, but the core idea is within scope."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive comments on the evaluation of perceptual claims. We address each major point below and indicate planned revisions.","responses":[{"response":"We agree the abstract phrasing is imprecise. The manuscript reports numerical joint-position and velocity errors on CMU Mocap sequences for the three ONNX precision variants, showing that error growth remains modest within the chosen distance bands, but contains no human perceptual study or rendering simulation. We will revise the abstract to state that numerical errors stay below a chosen threshold at the assigned ranges and explicitly note the absence of perceptual validation as a limitation of the current study.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the claim that 'evaluation on the CMU Mocap dataset provides initial evidence that each precision tier can be served at its assigned distance range with negligible perceptible degradation' supplies no quantitative error metrics (joint-position, velocity, or acceleration error), no description of how perceptual quality was assessed, and no human-subject or rendering-based validation that accounts for screen resolution, motion parallax, or distance-dependent visual acuity."},{"response":"The thresholds were derived from the point at which numerical error (joint position RMSE) begins to increase noticeably across the quantized models on the CMU test sequences. No separate perceptual calibration or in-game rendering study was performed. We will revise the evaluation section to (a) report the exact error values used to set the bands, (b) add a limitations paragraph stating that imperceptibility is an assumption based on numerical tolerance rather than measured human perception, and (c) outline how a future user study under game-like viewing conditions could be conducted.","revision_made":"yes","referee_comment":"[Evaluation] Evaluation section (referenced in Abstract): the distance thresholds that map FP32/FP16/INT8 tiers to ranges appear chosen to match observed numerical error tiers without an independent perceptual calibration step, human study, or simulation of in-game viewing conditions, leaving the weakest assumption (imperceptibility at assigned distances) unsupported."}],"tokens_in":1428,"tokens_out":442,"duration_ms":13634,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The new piece is treating inference precision tiers (FP32/FP16/INT8) as the LOD axis for a learned motion model in games, routing the ONNX variants by camera distance. That specific mapping is not in the Li et al. reference they build on, so the framing is fresh even if the underlying quantization trick is not.\n\nThey export the convolutional sequence-to-sequence checkpoint into three runtime variants and propose a distance selector at runtime. The CMU Mocap mention is meant to show that each tier can run in its assigned range without noticeable loss.\n\nThe problem is the evidence stays at the level of the abstract: no joint-position errors, no velocity errors, no description of how the distance thresholds were chosen, and no attempt to link numerical error to what a player actually sees at different screen distances or under motion. Standard mocap metrics do not automatically prove imperceptibility once you factor in visual acuity, animation artifacts, and screen resolution. The thresholds look like free parameters set to match the error tiers rather than calibrated against any perceptual data.\n\nThis is aimed at graphics engineers who already run learned animation and want to cut inference cost on distant NPCs. A reader already working on real-time AI LOD or quantization in games could pick up the idea quickly.\n\nIt is coherent on its own terms and engages the right prior work, so it clears the bar for a serious referee. I would send it to review only if the authors add the missing quantitative results and a short justification for the distance ranges; without those it stays a concept sketch.","headline":"The paper frames model quantization as a distance-based LOD switch for motion prediction but the evaluation supplies no numbers or perceptual checks to support the imperceptibility claim.","tokens_in":2269,"tokens_out":390,"would_cite":false,"duration_ms":14873,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Distance to the camera selects the numerical precision of ML models animating game characters.","keywords":["level of detail","quantization","motion prediction","game animation","inference optimization","character animation","machine learning"],"falsifier":"A controlled observation or player study in which viewers fail to detect quality differences between the assigned precision tiers at their respective distances, or conversely detect degradation inside the assigned range.","tokens_in":2601,"feed_emoji":"🎮","tokens_out":418,"duration_ms":14739,"temperature":0.7,"pith_summary":"The paper proposes AI Level of Detail, which adapts machine learning inference precision for motion prediction models according to each NPC's distance from the player camera. This mirrors classical geometry LOD but substitutes a lower-precision quantized model rather than a simpler mesh when the difference remains imperceptible. The authors export one trained convolutional sequence-to-sequence model into FP32, FP16, and INT8 variants and route them at runtime by distance, with evaluation on the CMU Mocap dataset indicating each tier can be used in its assigned range with negligible perceptible degradation. A sympathetic reader would care because game engines already expend substantial compute on animating multiple NPCs, and this technique offers a scalable way to reduce that cost without visible quality loss.","feed_headline":"Distance selects ML precision for NPC animations","feed_subtitle":"Lower-precision models for distant characters cut compute with no visible quality loss.","key_machinery":"A distance-based selector that routes between FP32, FP16, and INT8 per-tensor quantized ONNX Runtime variants of the same trained motion prediction model.","core_discovery":"Inference-time quantization can serve as the LOD axis for AI-driven character animation by routing lower-precision models to greater distances, with each precision tier assigned to a distance range such that resulting animation degradation stays imperceptible.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Distance routes quantized ML for NPC motion","AI LOD uses precision tiers by character distance","Quantized models serve as animation level of detail","Precision selection adapts ML to NPC distance"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Each precision tier can be assigned to a distance range such that the resulting animation degradation remains imperceptible to players.","fun_headline_variants_meta":{"raw":{"variants":["Distance routes quantized ML for NPC motion","AI LOD uses precision tiers by character distance","Quantized models serve as animation level of detail","Precision selection adapts ML to NPC distance"]},"model":"grok-4.3","cost_usd":0.003552,"raw_usage":{"total_tokens":1841,"prompt_tokens":625,"num_sources_used":0,"completion_tokens":52,"cost_in_usd_ticks":35524500,"prompt_tokens_details":{"text_tokens":625,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1164,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":625,"tokens_out":52,"duration_ms":7894,"temperature":1.0,"reasoning_tokens":1164,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T22:38:11.854458+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled observation or player study in which viewers fail to detect quality differences between the assigned precision tiers at their respective distances, or conversely detect degradation inside the assigned range.","supporting_citations":[],"review_version":1}