Game media and interactive narrative

Cinematics, playable stories, branching narrative, worlds and media assets

• articleWhat does 'frame pacing' mean, and why can a game at a steady 30 FPS still look uneven?
  Android's Swappy page defines frame pacing as synchronising the game loop with the display subsystem (not an average FPS) and illustrates uneven 49/16/33 ms frames from a nominal 30 FPS loop on a 60 Hz display; Godot separates jitter, stutter and input lag; Unity recommends hardware v-sync (vSyncCount > 0) for smooth pacing. Vendor taxonomies differ and do not map one-to-one.
• articleWhere is the Comfortaa font that the YarnSpinner-Unity notices mention, and why is it listed under 'Samples:'?
  The Comfortaa OFL notice in YarnSpinner-Unity has no Comfortaa-named file in the core tree at 90fb4c46 or in the YarnSpinner-Unity-Samples tree at 13f258df; the samples tree ships other fonts (four .ttf/.otf files) with their own licence files (producer).
• articleWhich binaries or source folders in YarnSpinner-Unity at 90fb4c46 correspond to each entry in its Third-Party Notices file?
  Each headed library notice (ANTLR4, Protocol Buffers, CsvHelper) has one Yarn.-prefixed DLL in Runtime/DLLs; three shipped analyser binaries in Editor/Analysis/DLLs (Microsoft.CodeAnalysis, Microsoft.CodeAnalysis.CSharp, System.Collections.Immutable, System.Reflection.Metadata) have no notice entry by name, and the notices file lists Microsoft/System names with no DLL in the tree.
• articleWhat third-party notices and licence statements ship with YarnSpinner-Unity at the pin used?
  YarnSpinner-Unity at 90fb4c46 ships a Third-Party Notices file separate from its MIT LICENSE.md, naming ANTLR4 and Protocol Buffers (BSD 3-Clause), CsvHelper (Apache 2.0), a Microsoft/System package list (MIT, .NET Foundation), Unity-SerializableDictionary (MIT) and, under Samples, the Comfortaa font (OFL 1.1, RFN Comfortaa).
• articleAt Godot 4.7.2-stable, how do resource UIDs (.uid sidecars, uid in .tscn/.tres headers and in .import) keep references stable, and what happens when files are moved, renamed or copied outside the editor?
  At Godot 4.7.2-stable (ed1daf0b), resources whose loader lacks custom UID support keep their UID in a '<file>.uid' sidecar, created only when none exists; text resources resolve ext_resource references UID-first and fall back to the stored path with an 'invalid UID ... using text path instead' warning; the UID registry is cached in uid_cache.bin under the project data path; the class docs state the sidecar rule.
• articleHow do the MPEG glTF extensions (MPEG_media, MPEG_buffer_circular, MPEG_accessor_timed, MPEG_texture_video, MPEG_scene_dynamic, MPEG_viewport_recommended) wire timed video, V3C and dynamic scene data into a glTF scene?
  The Khronos-hosted MPEG glTF extension READMEs (7cee0fcd) are based on ISO/IEC FDIS 23090-14 and define timed media through MPEG_media (alternatives with required mimeType, listed value video/mp4), a circular buffer (count default 2, minimum 2), a timed accessor that scenes with timed data shall use (suggested update rate default 25) and a video texture (format default RGB); each README states Copyright ISO/IEC 2022 with use subject to ISO's licence and IPR policy, not the repository licence.
• articleWhat do Xbox Accessibility Guideline 104 and the Game Accessibility Guidelines require for subtitles and captions in games beyond linear-video rules?
  Xbox Accessibility Guideline 104 asks for subtitles for all spoken content with speaker identification and spatial cues, independent on/off toggles for caption types such as dialogue, background dialogue and ambient sounds, lines of no more than about 40 characters, at most two lines, and text scalable to at least 200% of the default; the Game Accessibility Guidelines give 40 characters per line, two lines (three exceptionally) and a 46px-at-1080p minimum size.
• articleWhat preconditions does GGPO state for rollback netcode (determinism, fixed quanta, saveable state), and how does it detect violations?
  GGPO's developer guide (7ddadef8) requires a fully deterministic simulation, fully encapsulated and serializable game state, and the ability to load, save and step one frame without rendering; it recommends (not enforces) a fixed time quantum per frame and separating game state from render state, requires RNG seed and state in the game state, and offers a sync-test session that rolls back one frame every frame and errors on state differences. Its frame-delay thresholds (more than 1 frame noticeable in fighting games; 4-5 tolerable in board/puzzle games) are author judgement without cited measurements.
• articleHow do Ren'Py's save/load, rollback, default statement, after_load label and _version variable work together when a game script changes after players have saves?
  Ren'Py's docs (fbf45aa0) say engine-release save compatibility is not guaranteed, only variables changed after init (and defaulted ones) are saved, 'default' fills a variable missing at load, the after_load label can repair data (with renpy.block_rollback), and _version is set once at new-game start and must be updated by the game itself.
• articleHow does SugarCube 2 define what a save contains, what happens when a save from an older story version is loaded, and how can saves be versioned and upgraded?
  SugarCube 2 (docs at ba17b123) saves the state as of the last moment, which is created only on passage navigation; loading a save replaces the current state, so variables introduced in StoryInit after the save are undefined; save versioning is opt-in via Config.saves.version (no default) and is only useful with Save.onLoad upgrade handlers.
• articleWhat exactly is in an ink save-state JSON and which version numbers gate loading it?
  At ink 35c63e52 (v1.2.1) a save JSON records inkSaveVersion 10 and inkFormatVersion 21; loading rejects saves without inkSaveVersion or below 8, and compiled stories outside inkVersion 18-21 are rejected with a warning for in-range mismatches. The runtime-format document's example (inkVersion 10) is stale relative to the engine constant (21).
• articleWhat is the simplest test of whether a character agent refuses facts that the character could not know, and what numbers have been reported with it?
  Two papers use a multiple-choice test with an explicit refusal option. KBF-QA (arXiv 2606.25632v1): for a named character, each question has four answers plus 'I cannot answer this from my own knowledge'; items whose fact the character could witness have a normal gold answer, items it could not witness have the refusal as gold; 4,386 items over eight novels. CHARM (arXiv 2609.01352v1): separate 'is this outside the character's knowledge' yes/no questions and abstention-enabled questions, 40 characters, 2,748 questions. Reported: retrieve-everything RAG recalled 63.8 to 66.4 and refused only 8.7 to 11.3 on KBF-QA; the visibility-bounded ReverieMem refused 81.2; on CHARM GPT-4o recognised the boundary 91.3 percent of the time but still answered in 72.1 percent of cases.
• articleWhich licence terms are recorded for the Unity Recorder package components (source code, other package content, third-party GIF encoder) and what does the Unity Companion License v1.4 text say about use?
  The package's LICENSE.md (copyright 2022 Unity Technologies) says source code is under the Unity Companion License (UCL) "otherwise" under the Unity Package Distribution License (UPDL); Third Party Notices lists gif.ski as a commercial licence (Kornel Lesinski). The UCL v1.4 text served by unity.com on 2026-10-22 grants copyright rights only in connection with authoring/distributing content under a valid Unity Engine License, bars competitive analysis and competing products, and assigns derivative works of the Work to Unity. UPDL text and gif.ski terms were not read. No legal clearance is concluded.
• articleExactly how does the Matrix-Game GameWorld Score measure keyboard and mouse following, what do Oasis, MineWorld and Matrix-Game 1.0/2.0 score on it, and what does it need to run?
  GameWorld Score (Matrix-Game 1.0 paper and the GameWorldScore directory at commit c7a88bf) scores eight dimensions; the action ones come from an IDM trained on Minecraft (the paper says 90.6% keyboard accuracy, R2 0.97 for mouse) applied to generated clips. The code asks for 76 actions on 32 initial images, i.e. 2432 videos. Paper results (author-run): keyboard/mouse accuracy 0.77/0.56 Oasis, 0.86/0.64 MineWorld, 0.95/0.95 Matrix-Game 1.0. Matrix-Game 2.0 reuses the score on Minecraft (0.91/0.95) but says the action metric cannot be applied to wild scenes, so its comparison against YUME omits action accuracy. The benchmark downloads several checkpoints in pickle-capable formats and the repository carries conflicting licence signals.
• articleWhat evaluations of AI directors and emergent-narrative systems do the designers and open papers actually report, and how strong is the evidence that they improve player experience?
  The designer sources report outcomes, not measurements: Valve cites sales and awards for Left 4 Dead and asserts that a crude intensity estimate yields good pacing; no pacing data appear in the slides read, and RimWorld and Dwarf Fortress sources report none. A 2024 open paper says directly that no comparative data exist for the L4D director. Its own human-subject study of three directors in a small farming game (108 started, 32 usable) found that non-random directors beat random on time played and quests presented, with enjoyment lower only for random in the second session, yet its conclusion says players failed to perceive a difference. Valve's playtesting deck lists physiological arousal measures but contains no director-specific results. The evidence base for pacing controllers is therefore thin and mixed.
• articleIs WHAM's context 10 seconds or 1 second, and do the two statements count different units?
  They count different things, and the card contradicts itself. The Nature paper's methods say the 1.6B model's context is 5,560 tokens, 'equivalently 1 s or ten frames' (data downsampled to 10 Hz); its '10 s' is the length of the evaluated videos (consistency FVD uses 10-s clips). The Hugging Face card says 'Context length: 10 (observation, action) pairs / 5560 tokens' in the training section but 'Limited context length (10s)' under limitations. Ten pairs at 10 Hz is 1 s, so the card's '(10s)' does not match its own pair count; no Microsoft text explains the difference, so reading it as a slip for ten steps is an inference. Code defaults agree with ten frames: run_dreaming.py sets context_length 10.
• articleWhat structural statistics can I get from Twine, ink and Yarn Spinner without writing my own parser, and what do they not tell me?
  At the pinned versions: Twine's Details view reports characters, words, passages, links and broken links; inklecate's -s flag prints words, knots, stitches, functions, choices, gathers and diverts as JSON (source-level, not for compiled .json); `ysc graph` exports all nodes and links as DOT or mermaid. None reports number of endings, path counts, cycles or reconvergence; those need the exported graph plus your own computation.
• articleWhat does a drama manager consist of, and which ten desiderata does a survey of drama managers propose for judging one?
  Roberts and Isbell (2008) define a drama manager as a coordinator that tracks narrative progress and directs objects and agents toward a narrative or training goal. They say every approach rests on four parts: plot points (with precedence constraints), drama-manager actions, a model of player responses to those actions, and a model of authorial intent. They list ten desiderata (speed, coordination, replayability, authorial control, player autonomy, ease of authoring, adaptability, soundness, invisibility, measurability) and say some conflict, notably player autonomy and authorial control. The analysis is qualitative and, by the authors' own statement, speculative about what authors can actually exploit.
• articleHow do Ware, Young et al. define narrative conflict inside a causal-link plan (CPOCL), what are the seven conflict dimensions, what did the two human studies measure, and what can an outliner reuse?
  Ware and colleagues (IEEE T-CIAIG 2014) define conflict as a character forming a plan that is thwarted by another event, or would have been thwarted. In a plan with causal links and intention frames, a conflict is a threatened causal link in one character's subplan where at least one of the threatening or threatened steps is never executed; internal, external and environmental conflicts differ by who owns the threatening step (same character, another character, Fate/happening). Seven dimensions classify a conflict: participants, reason and duration (observable) and balance, directness, stakes and resolution (continuous 0-1 style measures). Study 1 (27 respondents, 23 finishing all three stories) found CPOCL's conflicts a superset of those readers reported; Study 2 (30 respondents) found the formula-predicted ordering was the top-ranked human ordering for balance, directness and resolution and fifth-ranked for stakes. The authors note durative actions and abstraction mismatches as li...
• articleHow does IPOCL (Riedl & Young 2010) represent character goals, motivations and causal links, what flaws and completeness conditions does it enforce, what are its stated limits, and what can an outliner reuse?
  IPOCL extends partial-order causal-link planning with 'frames of commitment': each frame records a character, the goal they pursue, the steps they perform for it, a final step that achieves it, and a motivating step that causes the character to intend it. A plan is complete only if all preconditions are established, causal threats resolved, and every non-happening step belongs to some frame; three added flaw types (open motivation, intent, intentional threat) drive the search. The article gives definitions, an algorithm and a complexity bound but no formal soundness or completeness theorem: a text search of the PDF for theorem/lemma/proof found none. Its evaluation used 32 CS undergraduates answering why-questions about POCL- versus IPOCL-generated text, judged against the QUEST model, with a significant effect for 'poor' answers and a p of about 0.058 for 'good' ones, without controlling narrative length. Stated limits: every intention is achieved, so repeated failure and hero-vers...
• articleWhat exactly does Microsoft's WHAM/Muse release offer for gameplay ideation, and how do the Nature paper and the model card differ?
  WHAM is a single-game model (Bleeding Edge), 300x180 at 10 Hz, with a 1-second context window, built from a user study of game creatives and evaluated on consistency (FVD), diversity (Wasserstein) and persistency (blinded human labels). The released 200M and 1.6B checkpoints are research-licensed (non-commercial, no distribution), not real time, and domain-bound. The model card's 'Limited context length (10s)' conflicts with the paper's 1 s of context; the card's own '10 pairs' agrees with the paper.
• articleWhat did Unreal Engine 5.1's documentation say about the OpenColorIO plugin and HDR display output for cinematics and virtual production?
  In Unreal 5.1 documentation, OCIO came through one OpenColorIO Configuration Asset used by the viewport/Play-in-Editor, Composure, Blueprints, nDisplay and Movie Render Queue, and the engine supported OCIO v1 configs only. Turning on the viewport OCIO display disables the Tone Curve, and MRQ's Color Output can apply an OCIO config and disable the Tone Curve, so whether written frames include the tone curve depends on that setting (see CA-B374 for MRQ file encodings). HDR display output was limited to 1000/2000-nit paths and D3D11 exclusive fullscreen. Confirm the engine release before relying on these limits.
• articleI want a ready-made ink or Yarn Spinner story to test a parser: which public ones exist and are they under a clear licence?
  The Intercept's README says The Intercept and ink are released under MIT, but a bundled Freesound sound-pack readme in the same repository marks listed sounds Attribution Noncommercial (CC BY-NC 3.0), so the repository-level MIT sentence does not cover all bundled media.
• articleWhich systems claim to steer a running world model or long video with text, how was that steering actually tested, and what did the tests show?
  Four measured protocols for text steering exist in 2025-2026 preprints, plus one system (Zing-0.5) that shows it only qualitatively, and each measures something different. Hunyuan-GameCraft-2 uses a VLM judge on six interaction dimensions over 100 test images against image-to-video baselines. Incantation uses blinded human ratings (three annotators, 400 clips) of whether an entity performed the prompted action. LongLive uses CLIP score per 10 s segment over 160 videos with six prompts each. WorldSonus compares a mid-stream sound prompt switch against a no-switch hold with identical visuals (audio only). Zing-0.5 measures only navigation (WBench) and shows text events as a qualitative session whose lasting consequences the authors say are not established. No source tests a text event in a released interactive model with persistence over later actions.
• articleHow does WorldMem store and retrieve earlier frames so a generated Minecraft or RealEstate10K world stays consistent after the camera leaves and returns, what revisit numbers does it report, and what does it cost?
  WorldMem keeps every generated frame as a token-level memory unit with a 5-number pose and a timestamp, picks up to a handful of units per step by field-of-view overlap and time difference (greedy, with a similarity filter), and lets the generating tokens cross-attend to them with pose/time embeddings. Its revisit tests compare regenerated frames against ground-truth frames with PSNR, LPIPS and reconstruction FID: 100 frames generated after a 600-frame memory bank on a custom Minecraft set, and five start=end trajectories over 100 RealEstate10K scenes. These are paper_result numbers on simulator poses; the paper's own tables disagree with each other and with the repository's expected-results block on some cells.
• articleHow does VMem index past views with surfels, what is its cycle-trajectory revisit test, and which parts of its release are open, gated or conflicting?
  VMem keeps the past views as images but indexes them through a coarse surfel model: each surfel stores which views saw it, new target cameras render the surfels to vote for the top-K most relevant past views, and an image-set generator (Stable Virtual Camera, SEVA) generates the next views from those. Its main contribution to evaluation is a cycle-trajectory protocol (go out, return along the same path, score every 10 frames of the return) with LPIPS/PSNR/SSIM plus rotation and translation distance of camera poses re-estimated from the generated frames. The tables contain rows that are identical across methods, which should be checked before reuse. Code is MIT-labelled but vendors a CC BY-NC-SA component, and the weights sit behind a Hugging Face access gate.
• articleIf I run the open Oasis 500M code, what resolution, frame rate, context length and action vector does it actually use, and which of those does the project text state?
  At commit f59deef2c019c212bd0c5a3a5b986a51f3701847 the open-oasis README states keyboard-conditioned autoregressive generation but no resolution, frame rate or context length; those come from the code. The code resizes prompts to 360x640, encodes with a ViT-VAE (patch 20) to an 18x32 latent grid, conditions on a 25-entry action vector (VPT-style keys plus two camera values), and slides a window of at most 32 frames (max_frames=32). The only '20' is a default for the saved video's fps, and the 20 fps real-time claim is on the joint blog about the full model (Decart's inference stack), not the open 500M weights. The weights repo is gated and was not opened.
• articleWhat does Epic's public documentation say about how the Derived Data Cache decides an entry is invalid or must be regenerated?
  The UE DDC documentation page is readable by plain GET (HTTP 200 on 2026-10-22) and describes cache hierarchy lookup/fill and disposability, not key composition or invalidation rules; the negative is about this page's content, not site readability.
• articleWhich engine docs claim, deny or condition deterministic simulation, and what do they say about fixed timesteps?
  Godot's docs (0914a4d4) state that physics in Godot, regardless of which physics engine Godot uses, is not deterministic.
• articleWhat do the ink JSON runtime format (v1.2.1) and the Twine 2 HTML/archive format preserve about line identity, localisation and variables?
  At ink 35c63e52 the JSON-format document example shows inkVersion 10 while the runtime defines inkVersionCurrent = 21; the JSON format document defines no localisation or line-identifier field in the sections read.
• articleWorldMark reports a 'Response Latency' for ten released interactive world models: what does it measure, is it compute or wall-clock latency, what are its values, and how do they relate to throughput figures?
  WorldMark v2 (Alaya Lab and others; not the model authors) scores 'Response Latency' as how soon the generated video's optical-flow speed reaches half its normal sustained speed after a command switch, normalised to 0-100 (higher is better). It is a content-response metric, not compute time: the paper reports no generation time or FPS, only that scoring one model on one split takes about one hour on two H200 GPUs excluding video generation. First-Person Real translational/rotational Response Latency: Matrix-Game 2.0 98.23/86.47, LingBot-World 86.56/99.20, HY-World 1.5 85.44/85.13, HY-GameCraft 1.0 90.97/51.99, Yume 1.5 40.66/77.11. Yume 1.5 has the lowest translational Response Latency here yet the highest Hz in WorldRoamBench, so the two measure different things. The v1 HTML has six models and no latency metric.
• articleWhat non-author evidence exists on the time from starting an interactive world model to the first usable frame, which parts of that cold-start time are separable (loading, compile or kernel warm-up, offload transfers), and which were fixed upstream?
  No non-author source times 'time to first displayed frame' as one span, but logs separate its parts. LingBot-World-Fast on 8xH100 (contributor's PRs): model construction about 129 s, prewarm about 7 s, first chunk about 7.2 s versus about 1.0 s steady; prewarm() was merged to take the first-chunk cost out of the user-facing window. HY-WorldPlay Wan-5B on 4xA100 (user): first four chunks 89-368 s, then about 3.4 s; a maintainer attributes it to torch.compile initialisation, and another maintainer warns the first run includes model loading. Matrix-Game-2: a user sees about a minute before the first response, tied by maintainers to the compiled VAE decoder. Repeated offload transfers add seconds per chunk even after start-up (HY-WorldPlay issue 53). Whether WorldRoamBench's 20-run averages exclude warm-up is not stated in the caption.
• articleWhich released interactive world models have a speed or latency figure measured by someone other than their authors, where is it, and which systems have none?
  Non-author timings exist for several systems but live in only four kinds of place: a benchmark's efficiency table (WorldRoamBench v5 times 10 open models on one H20), an efficiency paper's baseline row (QuantWM v3 times LingBot-World-v2 and HY-World 1.5 on A800 GPUs), profiler logs in GitHub pull requests (LingBot-World-Fast on 8xH100), and user bug reports (HY-WorldPlay, Hunyuan-GameCraft, Matrix-Game-2). No non-author key-press-to-frame measurement was found for any system, none was found for Matrix-Game 4.0 in the sources read, and the closed systems (Genie 3, Happy Oyster) have no timing in WorldRoamBench. WorldMark's 'latency' is an in-video response metric, not wall-clock time.
• articleWhat does Steam's Content Survey ask about AI-generated content, what changed in its wording, and what stays inside Steamworks behind login?
  Steamworks documentation (public page, captured 2026-10-22) says the Content Survey has a third section on generative AI that the developer must complete before review. It separates Pre-Generated content (shipped content created with AI tools during development) from Live-Generated content (created while the game runs), and for live-generated content the developer must describe guardrails against illegal output. Valve says it compares survey answers with the build and store page, and that the survey does not replace the content rules. Internet Archive captures show the Pre-Generated definition read "Any kind of content (art/code/sound/etc) created with the help of AI tools during development" at least until 2026-05-13, and by 2026-06-17 it was scoped to content "that ships with your game and is consumed by players", with an added sentence that efficiency gains from built-in AI tools are not the focus. A secondary report dated a rewrite to 16 January 2026; the official page captured ...
• articleWhich documented techniques keep an LLM from giving a character information that character should not have, and what effect did each show?
  Six techniques have measured effects, each in a different setting, so they are not comparable head-to-head. (1) Two-pass perspective filtering (SimToM): GPT-3.5-Turbo ToMi false-belief 81.0 (+13.8 over 0-shot); single-prompt version loses the gain; human-annotated perspectives reach 96/96/82/98; mixed on OpenToM (candidate 04). (2) Per-character belief graphs (SymbolicToM): GPT-3-Davinci on self-built structure sets 20/25/39 to 92/91/90; needs chronological stories and suffers from extraction errors. (3) Visibility-tagged fact memory with scene anchoring (ReverieMem): refusal 81.2 versus 8.7 to 11.3 for retrieve-everything RAG; ablations show both layers matter; one speaker only. (4) Separate agent contexts versus an omniscient script (Zhou et al.): fewer early leaks (0.39 vs 0.13) but far fewer deals (30 vs 94 percent). (5) Awareness reminder (SimpleToM): behaviour prediction above 80 for three models, judgment low. (6) Key Facts Injection: 73 percent fewer 'known-information forge...
• articleHow have unearned character knowledge or leaked information been detected and rated in LLM-generated stories and simulated conversations?
  Two partial measurements exist in the sources inspected; no source measured character-to-character secret leakage in generated fiction with human raters (bounded negative). (1) ConStory-Checker (arXiv 2603.05890v1) is a four-stage LLM judge with a 'Knowledge Contradictions' sub-type (characters showing knowledge or skills beyond their background, including information they could not have accessed given their background), with quoted evidence and positions; validated on 1,000 planted errors: precision 0.884, recall 0.550, F1 0.678 against human experts at F1 0.281. No sub-type rate is reported in the text read. (2) Zhou et al. (arXiv 2403.05020v4) use the early position of a name in a find-the-mutual-friend dialogue as a leakage proxy: 0.13 for an omniscient single-LLM script versus 0.39 with separate agents (450 simulations per model and mode).
• articleIf an LLM correctly says a character does not know a secret, will it then predict that the character acts as if they do not know?
  SimpleToM (arXiv 2410.13648v2) uses two-sentence everyday stories (a hidden fact; a person acting without knowing it) and asks three questions: is X aware (mental state), what will X do (behaviour), was X's action reasonable (judgment). Models were near ceiling on awareness (GPT-4o 95.6, Claude-3.5-Sonnet 97.9) but far lower on behaviour (49.5, 67.0) and judgment (15.3, 24.9) where chance is 50; o1-preview reached 84.1 on behaviour. The failure in the authors' example is predicting the observer acts on the hidden fact. Reminders of the awareness answer lifted behaviour above 80 for three models but judgment stayed low.
• articleWhen an LLM reads a conversation where a character was absent for part of it, how often does it credit that character with the missed information, and how consistent are its answers across ways of asking?
  FANToM (Kim et al., arXiv 2310.15421v3) builds 256 multi-party conversations in which a character leaves and rejoins, then asks the same underlying question six ways (free-form belief, multiple-choice belief, answerability list and yes/no, information-access list and yes/no). The model reads everything, so an answer containing what the absent character missed is the failure. In the 2023 models tested, success on one format did not carry to the others: the best chain-of-thought GPT-4 run scored 26.6 on 'all question types' against 87.5 for humans, and the dominant list error was including characters who were unaware. Results are for 2023 models.
• articleWhich story-based theory-of-mind benchmarks test which character knows or believes what, what does each actually ask, and how large is each?
  Five public benchmarks are inspected. ToMi: Sally-Anne object-location stories (999 narratives per OpenToM's table; paper body not read). Hi-ToM: generated stories with five agents and one question each at orders 0 to 4, optional public or private deceptive statements. OpenToM: 696 narratives with 23 first- and second-order questions each (location, multi-hop container/accessibility, attitude). FANToM: 256 conversations in which a character leaves and rejoins; belief, answerability and information-access questions. SimpleToM: two-sentence everyday stories, 1147 per category, with awareness, behaviour and judgment questions. All are template-built or LLM-generated; none uses published fiction.
• articleDo LLM-played characters keep their voice over a long dialogue, how is drift measured, and what do the authors say it converges to?
  Three open-access papers measure drift without human raters in the main loop. Li et al. (arXiv 2402.10962v4) report instruction drift within eight rounds of two-chatbot self-chat using code-based probes. Luz de Araujo et al. (arXiv 2512.12775v2, EACL 2026 per abstract) run over 100-round dialogues with eight RoleBench characters across seven models and report that fidelity (knowledge, style, in-character consistency) falls and persona behaviour reverts toward a no-persona baseline. Abdulhai et al. (arXiv 2511.00222v1) show that a line-to-line consistency metric can rise while prompt consistency falls. The judge validations are small or single-rater.
• articleHow was PersonaGym's PersonaScore validated against human judgement, and what does its Linguistic Habits task say about voice?
  PersonaGym (arXiv 2407.18416v5, 5 Sep 2025) scores persona agents on five tasks with an ensemble of two LLM judges using 1-5 rubrics. A human study of five experts rated 1,500 responses (100 personas, three models) and reports Fleiss kappa 0.71 among annotators and average Spearman 75.1% and Kendall 62.73% between PersonaScore and humans. Linguistic Habits was the hardest task. Personas are described people (for example professions), not fictional characters, and the text and the correlation table disagree on one Kendall claim.
• articleWhich English-language role-play or persona benchmarks exist, and what does each one actually score (is character-to-character voice distinctiveness among the scored things)?
  Four benchmark-style resources were read. RoleBench scores reference overlap (ROUGE-L) and judge/human win rates for single-turn answers in a role; PersonaGym scores five rubric tasks per demographic persona with an LLM-judge ensemble; PersonaEval scores whether a judge can name the speaker in human-written dialogue; the Speaker Verification work is the only one read that defines a cross-character measure (Distinction Score) on generated lines. CharacterEval is Chinese-only on its abstract page. None of the four reports a human-rated test of whether generated characters sound different from each other.
• articleIf I track inventory, location, character knowledge and consequences across many turns of an interactive story, which failure modes have these two papers measured and which are only proposals?
  Documented use (measured in the papers, on non-story environments): predicted-state quality drops with horizon when predictions are fed back (S2), a model that predicts from raw history can claim an object is unavailable or return the previous room (S2 Appendix G examples, selected), a learned simulator can let weak agents "win" it could not win for real (S1 Table 5), and one world model lost room connectivity in a looping TextWorld run (S1 Appendix H, one trajectory). Not measured: character knowledge or belief differences between characters, NPC dialogue, long narrative sessions (the longest horizon is 20 actions in Jericho; trajectories average 28.6-52 steps), or any per-category error rate for inventory. Everything about running such a tracker for authored fiction is a hypothesis.
• articleWhat exactly did arXiv 2610.06637 (v1) define and measure for long-horizon textual world modelling: environments, horizons, models, metrics, headline numbers, ablations, limits, release?
  arXiv 2610.06637 (Wang, Gao, Yao, Dong, Walia, Das; UIC and Intuit AI Research; v1 submitted 2026-10-05 16:25:44 UTC; CC BY 4.0) fine-tunes Qwen3-4B-Instruct-2507 world models that keep a JSON "textual state" and predict a whole k-action sequence in one call as sparse set/remove edits that code replays (latent-SR). It evaluates horizons 1-10 (ScienceWorld, CEO-Bench) and 1-20 (Jericho) with a fixed-target protocol against raw-history recursive/direct baselines and GPT-5.2. latent-SR has the best averages on every metric (ScienceWorld Fact-F1 92.44 vs 88.94 for obs-direct; CEO-Bench SE-score 87.35 vs 81.85; Jericho BLEU 24.5 and Score MAE 1.728), with gains growing at long horizons, but obs-recursive is better on Jericho at h1 and h5. The abstract's "nearly 30%" Jericho error drop holds against obs-direct at h10 (28.9%) but not against obs-recursive (12.2%), which is the lower-error raw-history baseline there. No code, data or weights release is stated.
• articleWhich dimensions of VBench-2.0, WorldSimBench and EWMBench would an agent actually use to screen a video model or world model for previs or game prototyping, and which gaps do none of them cover?
  Synthesis: for previs clips, the usable pieces are VBench-2.0's five custom-input dimensions (Human Anatomy, Identity, Clothes, Multi-View Consistency, Diversity) as cheap screens, and, only at model level on the fixed prompt suite, Camera Motion, Instance Preservation and Motion Rationality; do not expect Complex Plot or mid-clip attribute/position changes to work (authors report near-floor scores). For game prototyping none of the three measures revisit persistence, input latency, long-horizon drift or user-controllable rollouts; WorldSimBench's Minecraft closed loop is the nearest idea but its code and evaluator were not found, and EWMBench is robot-manipulation with fixed viewpoint and a robot-specific judge prompt. This is a proposed screening checklist, not a tested protocol.
• articleWho owns 3D models generated with Meshy and Hyper3D Rodin, what do they cost, and what are the retention/training terms?
  Meshy Terms (last updated 2026-09-19): on the free plan Meshy owns all rights in outputs and makes them available under CC BY 4.0 with credit to Meshy; on paid plans customers own their outputs (to the extent law allows) and may keep content private; outputs may carry AI-identifying metadata/watermarks that users must not remove; non-Enterprise inputs/outputs may be used to train or improve services; API outputs (non-Enterprise) are deleted after 3 days, so download promptly. Meshy API pricing is in prepaid credits per request (credit-to-USD rate not on that page): e.g. Image/Text-to-3D preview 20 credits on meshy-7.1/meshy-6 (5 on meshy-6-lite or the smart-topology t2/t2.5 models), mesh with 2K-4K textures 30, 8K 35; refine/retexture 10; auto-rigging 5; animation 3 per action; remesh 5; smart-topology polycount 100-15,000 (t2) or 100-30,000 (t2.5). Hyper3D (Rodin) pricing page (rendered 2026-10-23): Free $0 plan pays per confirmed result at $1.5 per direct credit, with export from ...
• articleBesides retrieving raw frames, how do a fused point-map spatial memory (SPMem) and an addressable compressed KV-cache memory (WorldTrace) keep revisits consistent, and what revisit tests and costs do they report?
  SPMem (Wu et al.) conditions a CogVideoX-based generator on three memories: recent frames, a static point cloud fused by TSDF from predicted point maps, and a few historical reference frames; its revisit test pairs frames at the same camera location in forward and reversed trajectories. WorldTrace (Wu, Elflein et al., NVIDIA) is training-free: it argues stored KV entries become unreadable when rotary position offsets leave the training range, assigns each compressed summary a fixed in-distribution position, and keeps verbatim 'landmark' scene entries; its LoopBench revisits a starting scene along ABA, ABCA and ABCDA loops and scores the return frame against the earlier generated frame with CLIP similarity. Both report paper_result numbers only; SPMem ships Apache-labelled code and checkpoints, WorldTrace code was not found.
• articleHow does Context-as-Memory pick which earlier frames to condition on, how does it test scene consistency, and what can an outside agent obtain from it?
  Context-as-Memory (CaM) stores past frames unprocessed, concatenates retrieved ones with the frames being predicted along the frame axis, and retrieves by camera-trajectory field-of-view overlap (then drops adjacent duplicates and optionally adds far-in-space/time frames). It evaluates with PSNR/LPIPS against ground truth and a stronger 'history context' comparison where a camera rotates n degrees and returns and the new frames are compared with earlier generated frames. Reported on an Unreal Engine 5 set of 100 long videos with a 1B internal base model; code was 'Coming Soon' on the project page and the GitHub path was not found, while the dataset is on Hugging Face without a licence field.
• articleWhat did each of these open interactive world models actually evaluate (metrics, baselines, raters, test sets), and what do the evaluations not support?
  The evaluations are author-run and use different protocols. Matrix-Game 1.0 uses its own GameWorld Score on Minecraft plus a double-blind human comparison against Oasis and MineWorld. Matrix-Game 2.0 reports the same score against YUME on wild scenes and is lower on motion smoothness, object consistency and scenario consistency. Matrix-Game 4.0's report and card give latency and throughput but no image-quality metric, and the card says it claims no benchmark. Hunyuan-GameCraft reports FVD, relative pose error, image quality and a 30-evaluator ranking on 150 images and 12 action signals, compared with camera-control baselines and Matrix-Game 1.0. LingBot-World 1 uses VBench on 100 generated videos over 30 s; LingBot 2.0 relies on comparison figures and an hour-long session. None measures persistence or revisit consistency quantitatively.
• articleWhat is the architecture of the "Generative Agents" LLM characters (memory, reflection, planning), stated plainly?
  Park et al. (arXiv 2304.03442, pinned v2, UIST 2023) describe an LLM wrapper with a memory stream of natural-language observations, a retrieval function that scores memories by recency, importance and relevance, periodic reflection that writes higher-level memories, and planning that turns the agent's summary and memories into daily plans. It was instantiated with gpt3.5-turbo in a 25-agent sandbox. These are the authors' reported implementation values, not a production game recipe.

More