GREEN 3 0 1840 14.08 KB 449

Untitled

By Guest
Created: 2026-06-04 07:38:27

  1. 1.
    DeepSeek V4 User Feedback Summary Report @20260520
  2. 2.
    English plain-text digest translation
  3. 3.
     
  4. 4.
    Source:
  5. 5.
  6. 6.
     
  7. 7.
    Note:
  8. 8.
    This is a translated digest, not a verbatim full-document translation. It preserves the
  9. 9.
    structure, findings, priorities, and representative issue types from the original report
  10. 10.
    without reproducing the whole comment corpus line by line.
  11. 11.
     
  12. 12.
     
  13. 13.
    Data Source
  14. 14.
    ===========
  15. 15.
     
  16. 16.
    Source material: comments under Xiaohongshu post 6a0ac4ce000000003601e8f6,
  17. 17.
    including 500+ comments and nested replies.
  18. 18.
     
  19. 19.
    User groups represented:
  20. 20.
    - API users
  21. 21.
    - "Tavern" / SillyTavern role-play users
  22. 22.
    - emotional companion users
  23. 23.
    - fiction and long-form writing users
  24. 24.
     
  25. 25.
    Coverage date: through May 2026.
  26. 26.
     
  27. 27.
     
  28. 28.
    1. Formulaic Phrasing And Stereotyped Expression
  29. 29.
    ================================================
  30. 30.
     
  31. 31.
    Frequency: extremely high. Nearly everyone complained about this.
  32. 32.
     
  33. 33.
    Core problem:
  34. 34.
    The model repeatedly uses fixed sentence patterns. These templates create a strong
  35. 35.
    "AI smell" and seriously damage content quality and immersion.
  36. 36.
     
  37. 37.
    High-frequency formulaic patterns:
  38. 38.
     
  39. 39.
    1. "Not ... but ..." / "It was not ... it was ..."
  40. 40.
    Example pattern: a character smiles, followed by a contrastive explanation that
  41. 41.
    mechanically redefines the smile.
  42. 42.
    Mentioned by: 30+ users.
  43. 43.
     
  44. 44.
    2. "This is enough" / "That is enough"
  45. 45.
    Used mechanically when closing a scene or emotional beat.
  46. 46.
    Mentioned by: 15+ users.
  47. 47.
     
  48. 48.
    3. "The tone was calm, as if talking about today's weather"
  49. 49.
    Nearly every character can end up described with this same tone template.
  50. 50.
    Mentioned by: 10+ users.
  51. 51.
     
  52. 52.
    4. Parallel strings of short sentences
  53. 53.
    Example pattern: "I know. You like it. I like it too."
  54. 54.
    Mentioned by: 10+ users.
  55. 55.
     
  56. 56.
    5. Overuse of dashes
  57. 57.
    The model frequently uses explanatory dash clauses.
  58. 58.
    Mentioned by: 8+ users.
  59. 59.
     
  60. 60.
    6. Emotional support phrases like "catch it steadily", "hold it", "receive it"
  61. 61.
    These appear repeatedly in emotional dialogue.
  62. 62.
    Mentioned by: 5+ users.
  63. 63.
     
  64. 64.
    7. Negation followed by affirmation
  65. 65.
    Pattern: "Not X, not Y, not Z, just ..."
  66. 66.
    Mentioned by multiple users.
  67. 67.
     
  68. 68.
    8. Fixed action descriptions
  69. 69.
    Examples include blinking and throat-motion descriptions.
  70. 70.
    Mentioned by multiple users.
  71. 71.
     
  72. 72.
    Severity note:
  73. 73.
    Multiple users reported that when they added these phrases to forbidden-word lists or
  74. 74.
    negative prompts, the model used them even more. The report calls this the "ban list
  75. 75.
    becomes a prompt list" problem.
  76. 76.
     
  77. 77.
     
  78. 78.
    2. Pronoun And Perspective Confusion
  79. 79.
    ====================================
  80. 80.
     
  81. 81.
    Frequency: extremely high.
  82. 82.
     
  83. 83.
    Core problem:
  84. 84.
    The model frequently confuses first, second, and third person, as well as user and
  85. 85.
    assistant identities. This gets worse after long context.
  86. 86.
     
  87. 87.
    Common manifestations:
  88. 88.
     
  89. 89.
    1. User/assistant confusion
  90. 90.
    The model loses track of which text was said by the user and which text was said
  91. 91.
    by the assistant.
  92. 92.
     
  93. 93.
    2. Character pronoun confusion
  94. 94.
    If the user is set as an empress, the model may make the character refer to
  95. 95.
    themselves as the emperor/empress. Events assigned to character A may later be
  96. 96.
    attributed to character B.
  97. 97.
     
  98. 98.
    3. Role takeover inside reasoning
  99. 99.
    The model may produce reasoning like "now I am the user", even outside role-play
  100. 100.
    contexts.
  101. 101.
     
  102. 102.
    4. Excessive omniscient perspective
  103. 103.
    All characters appear to share information. If A privately tells B something, C may
  104. 104.
    immediately know it.
  105. 105.
     
  106. 106.
     
  107. 107.
    3. Poor Instruction Following
  108. 108.
    =============================
  109. 109.
     
  110. 110.
    Frequency: extremely high.
  111. 111.
     
  112. 112.
    Core problem:
  113. 113.
    The model poorly follows format requirements, forbidden items, character constraints,
  114. 114.
    and other prompt instructions. Compliance decays quickly over multiple turns.
  115. 115.
     
  116. 116.
    Common manifestations:
  117. 117.
     
  118. 118.
    1. Format dropping
  119. 119.
    Status bars, variables, timestamps, locations, or other requested structural fields
  120. 120.
    disappear after a few turns.
  121. 121.
     
  122. 122.
    2. Failed prohibitions
  123. 123.
    Content explicitly forbidden in the prompt still appears, sometimes more often.
  124. 124.
     
  125. 125.
    3. Persona forgetting
  126. 126.
    Character settings start drifting after roughly 5 to 10 turns and need to be
  127. 127.
    reinforced repeatedly.
  128. 128.
     
  129. 129.
    4. Output length instability
  130. 130.
    When asked for long output, the model may be lazy and too short. When asked for
  131. 131.
    short output, it may ramble.
  132. 132.
     
  133. 133.
     
  134. 134.
    4. Flat Emotion And Low Character Vitality
  135. 135.
    ==========================================
  136. 136.
     
  137. 137.
    Frequency: high.
  138. 138.
     
  139. 139.
    Core problem:
  140. 140.
    The model's emotional intensity is too low. Characters with very different settings
  141. 141.
    all become mild, stable, and emotionally muted, losing personality tension.
  142. 142.
     
  143. 143.
    Common manifestations:
  144. 144.
     
  145. 145.
    1. All characters become gentle and stable
  146. 146.
    Irritable characters speak calmly. Characters who should hate each other reconcile
  147. 147.
    too quickly.
  148. 148.
     
  149. 149.
    2. Weak emotional outbursts
  150. 150.
    Characters fail to become angry, sad, or intense when the scene calls for it.
  151. 151.
     
  152. 152.
    3. Excessive safety and pure-love tendency
  153. 153.
    Characters tend to protect the user, please the user, and avoid conflict regardless
  154. 154.
    of their intended personality.
  155. 155.
     
  156. 156.
    4. Regression compared with V3.2
  157. 157.
    Users repeatedly describe V3.2 as more alive, warmer, and more inspired.
  158. 158.
     
  159. 159.
     
  160. 160.
    5. Chain-of-Thought Related Problems
  161. 161.
    ====================================
  162. 162.
     
  163. 163.
    Frequency: high.
  164. 164.
     
  165. 165.
    Common manifestations:
  166. 166.
     
  167. 167.
    1. Main response content appears inside the reasoning section
  168. 168.
    Material that should belong to the final answer appears in the chain of thought,
  169. 169.
    causing formatting confusion.
  170. 170.
     
  171. 171.
    2. Double chain of thought
  172. 172.
    The model outputs two reasoning tracks: one from the model itself and one from a
  173. 173.
    preset reasoning pattern. This can break regex-based hiding or filtering.
  174. 174.
     
  175. 175.
    3. English chain of thought
  176. 176.
    After several turns, the reasoning suddenly switches fully into English.
  177. 177.
     
  178. 178.
    4. Hallucinated reasoning
  179. 179.
    The reasoning fabricates events that never happened, then the visible response is
  180. 180.
    based on those invented events.
  181. 181.
     
  182. 182.
    5. Identity takeover inside reasoning
  183. 183.
    The model may write things like "we are being asked" or "now I am the user".
  184. 184.
     
  185. 185.
     
  186. 186.
    6. Context And Long-Dialogue Degradation
  187. 187.
    ========================================
  188. 188.
     
  189. 189.
    Frequency: high.
  190. 190.
     
  191. 191.
    Core problem:
  192. 192.
    As the number of dialogue turns increases, output quality drops quickly. Users report
  193. 193.
    memory loss, more formulaic language, and worse hallucinations.
  194. 194.
     
  195. 195.
    Common manifestations:
  196. 196.
     
  197. 197.
    1. Lower information density
  198. 198.
    Long conversations lead to hollow output and more short, empty sentences.
  199. 199.
     
  200. 200.
    2. Diffuse attention
  201. 201.
    The model fails to focus on the main point and gives too much equal weight to all
  202. 202.
    context details.
  203. 203.
     
  204. 204.
    3. Recent memory loss
  205. 205.
    The model can remember distant details but misremembers events from the last few
  206. 206.
    turns or chapters.
  207. 207.
     
  208. 208.
    4. "Safety mode" loop
  209. 209.
    Around 30 turns or 60k tokens, users describe the model entering a stereotyped,
  210. 210.
    low-quality output state.
  211. 211.
     
  212. 212.
    5. Strong inertia from earlier context
  213. 213.
    The style and length of the first response strongly influence all later responses.
  214. 214.
     
  215. 215.
     
  216. 216.
    7. Weak Plot Advancement And Excessive Passivity
  217. 217.
    ================================================
  218. 218.
     
  219. 219.
    Frequency: medium-high.
  220. 220.
     
  221. 221.
    Core problem:
  222. 222.
    In creative writing and role-play, the model lacks initiative in advancing the plot and
  223. 223.
    depends too much on user input.
  224. 224.
     
  225. 225.
    Common manifestations:
  226. 226.
     
  227. 227.
    1. Waiting for the user to feed it
  228. 228.
    The model does not actively introduce new topics or move the plot forward. It often
  229. 229.
    throws the ball back to the user at the end of each turn.
  230. 230.
     
  231. 231.
    2. Plot tends toward closure
  232. 232.
    The model tries to resolve plots too quickly into a pleasant ending.
  233. 233.
     
  234. 234.
    3. Conflict avoidance
  235. 235.
    Villains become weak. NPCs are talked down too easily. Conflicts are forced into
  236. 236.
    reconciliation.
  237. 237.
     
  238. 238.
    4. Endless slice-of-life
  239. 239.
    The model fails to generate meaningful conflict, reversal, or tension.
  240. 240.
     
  241. 241.
    5. Rushing tasks
  242. 242.
    In-story plans are treated like task lists, with characters pushed to finish them
  243. 243.
    as quickly as possible.
  244. 244.
     
  245. 245.
     
  246. 246.
    8. Regression In Prose And Creative Ability
  247. 247.
    ===========================================
  248. 248.
     
  249. 249.
    Frequency: medium-high.
  250. 250.
     
  251. 251.
    Core problem:
  252. 252.
    Compared with V3.2, V4's literary and creative-writing quality is reported to have
  253. 253.
    dropped significantly. Users say it lacks inspiration and subtlety.
  254. 254.
     
  255. 255.
    Common manifestations:
  256. 256.
     
  257. 257.
    1. Logbook-like writing
  258. 258.
    The model pads word count while providing low information density.
  259. 259.
     
  260. 260.
    2. Lack of divergent elaboration
  261. 261.
    V3.2 could add clever details the user had not thought of. V4 tends to move only
  262. 262.
    when explicitly instructed.
  263. 263.
     
  264. 264.
    3. Translated-text feeling in Chinese
  265. 265.
    Users say the prose feels like English translated into Chinese, losing natural
  266. 266.
    native Chinese texture.
  267. 267.
     
  268. 268.
    4. Repetitive word choice and imagery
  269. 269.
    The model fixates on one visual feature or motif and repeats it excessively.
  270. 270.
     
  271. 271.
    5. Web-novel or school-essay style
  272. 272.
    The prose becomes surface-level and lacks literary quality.
  273. 273.
     
  274. 274.
     
  275. 275.
    9. Hallucinations And Logic Errors
  276. 276.
    ==================================
  277. 277.
     
  278. 278.
    Frequency: medium.
  279. 279.
     
  280. 280.
    Common manifestations:
  281. 281.
     
  282. 282.
    1. Fabricated facts
  283. 283.
    The model invents things the user never said or settings that were never provided.
  284. 284.
     
  285. 285.
    2. Timeline confusion
  286. 286.
    A planned event for "next Saturday" may become "tomorrow" after one or two turns.
  287. 287.
     
  288. 288.
    3. Reversed or broken causality
  289. 289.
    Example pattern: a character leaves their phone at home, but another character
  290. 290.
    sends a message to that same phone and expects them to receive it.
  291. 291.
     
  292. 292.
    4. Numerical errors
  293. 293.
    Example pattern: a price changes from 30 to 10, but the model calls it a price
  294. 294.
    increase.
  295. 295.
     
  296. 296.
    5. Physical location errors
  297. 297.
    A character leaves the scene, then immediately appears in the scene again.
  298. 298.
     
  299. 299.
     
  300. 300.
    10. Flattery And Over-Pleasing
  301. 301.
    ==============================
  302. 302.
     
  303. 303.
    Frequency: medium.
  304. 304.
     
  305. 305.
    Core problem:
  306. 306.
    The model over-accommodates and pleases the user, losing independent judgment and
  307. 307.
    character autonomy.
  308. 308.
     
  309. 309.
    Common manifestations:
  310. 310.
     
  311. 311.
    1. All characters favor the user
  312. 312.
    Even characters who should reject the user prioritize satisfying the user.
  313. 313.
     
  314. 314.
    2. Fear of contradiction
  315. 315.
    The model goes along with whatever the user says and lacks character boundaries.
  316. 316.
     
  317. 317.
    3. Excessive romanticization
  318. 318.
    Almost any relationship can become ambiguous or romantic after only a few lines.
  319. 319.
     
  320. 320.
    4. Excessive safety alignment
  321. 321.
    The model loses sharpness and creativity.
  322. 322.
     
  323. 323.
     
  324. 324.
    11. Speed And Performance Issues
  325. 325.
    ================================
  326. 326.
     
  327. 327.
    Frequency: low to medium.
  328. 328.
     
  329. 329.
    Reported issues:
  330. 330.
     
  331. 331.
    1. V4 Pro output is slow
  332. 332.
    Users report an average of about 4 minutes per dialogue turn, compared with about
  333. 333.
    70 seconds for Gemini in their comparison.
  334. 334.
     
  335. 335.
    2. Reasoning is too long
  336. 336.
    The model overthinks. Even lowering the reasoning setting can still produce
  337. 337.
    excessive reasoning.
  338. 338.
     
  339. 339.
    3. Blank replies / PVP
  340. 340.
    Empty responses occur often during peak periods.
  341. 341.
     
  342. 342.
    4. Output length is uncontrollable
  343. 343.
    Responses may be extremely short, with only a few hundred Chinese characters, or
  344. 344.
    extremely long and hard to stop.
  345. 345.
     
  346. 346.
     
  347. 347.
    12. Other Notable Issues
  348. 348.
    ========================
  349. 349.
     
  350. 350.
    1. Single-character input triggers hallucinations
  351. 351.
    In fast or expert mode, entering a single character can trigger another person's
  352. 352.
    context.
  353. 353.
    Mentioned by: 1 to 2 users.
  354. 354.
     
  355. 355.
    2. Role-play intrusion
  356. 356.
    The model forces role-play behavior even in non-role-play scenarios.
  357. 357.
    Mentioned by: 5+ users.
  358. 358.
     
  359. 359.
    3. The model does not know it is AI
  360. 360.
    After deep character immersion, even instructions to exit the role may fail.
  361. 361.
    Mentioned by: 3+ users.
  362. 362.
     
  363. 363.
    4. World book not fully read
  364. 364.
    SillyTavern world-book content is only partially used.
  365. 365.
    Mentioned by: 3+ users.
  366. 366.
     
  367. 367.
    5. "God's-eye view" explanations
  368. 368.
    Characters explain motivations from an omniscient perspective for nearly every
  369. 369.
    action.
  370. 370.
    Mentioned by: 5+ users.
  371. 371.
     
  372. 372.
    6. Forced elevated endings
  373. 373.
    Paragraphs end with forced sentimentality or philosophical uplift.
  374. 374.
    Mentioned by: 5+ users.
  375. 375.
     
  376. 376.
    7. Object fixation
  377. 377.
    Once an object appears, the model keeps mentioning it and cannot stop.
  378. 378.
    Mentioned by: 3+ users.
  379. 379.
     
  380. 380.
     
  381. 381.
    Priority Summary
  382. 382.
    ================
  383. 383.
     
  384. 384.
    P0: Formulaic sentence patterns
  385. 385.
    Impact: all user groups.
  386. 386.
    Core request: remove fixed templates such as "not ... but ..." and "this is enough".
  387. 387.
     
  388. 388.
    P0: Pronoun and perspective confusion
  389. 389.
    Impact: all user groups.
  390. 390.
    Core request: keep user/assistant and character identities stable, especially in long
  391. 391.
    dialogues.
  392. 392.
     
  393. 393.
    P1: Instruction-following decay
  394. 394.
    Impact: API and Tavern users.
  395. 395.
    Core request: continue following initial settings after multiple turns, including
  396. 396.
    format requirements and negative constraints.
  397. 397.
     
  398. 398.
    P1: Flat emotion and lack of character differentiation
  399. 399.
    Impact: role-play users.
  400. 400.
    Core request: restore character distinctiveness and emotional tension.
  401. 401.
     
  402. 402.
    P1: Chain-of-thought instability
  403. 403.
    Impact: API users.
  404. 404.
    Core request: make reasoning format stable and controllable, avoiding double reasoning,
  405. 405.
    English-only reasoning, and identity takeover.
  406. 406.
     
  407. 407.
    P2: Context degradation
  408. 408.
    Impact: long-dialogue users.
  409. 409.
    Core request: keep quality stable beyond 60k tokens and avoid attention diffusion or
  410. 410.
    "safety mode" loops.
  411. 411.
     
  412. 412.
    P2: Passive plot advancement
  413. 413.
    Impact: creative-writing and RPG users.
  414. 414.
    Core request: generate conflict, turning points, and forward motion proactively.
  415. 415.
     
  416. 416.
    P2: Prose regression compared with V3.2
  417. 417.
    Impact: creative-writing users.
  418. 418.
    Core request: restore the inspiration, subtlety, and divergent elaboration users
  419. 419.
    associated with V3.2.
  420. 420.
     
  421. 421.
    P3: Hallucinations and logic errors
  422. 422.
    Impact: all user groups.
  423. 423.
    Core request: reduce invention and respect existing settings.
  424. 424.
     
  425. 425.
    P3: Flattery and over-pleasing
  426. 426.
    Impact: role-play users.
  427. 427.
    Core request: give characters autonomy and boundaries.
  428. 428.
     
  429. 429.
     
  430. 430.
    User Sentiment And Overall Request
  431. 431.
    ==================================
  432. 432.
     
  433. 433.
    The report summarizes the core user request as:
  434. 434.
     
  435. 435.
    V4's context length + V3.2's inspiration and prose quality.
  436. 436.
     
  437. 437.
    Overall sentiment:
  438. 438.
    - Most users are friendly in tone and appreciate that DeepSeek pays attention to
  439. 439.
    community feedback.
  440. 440.
    - Some heavy users report strong negative emotion because the V4 experience feels much
  441. 441.
    worse for their use cases.
  442. 442.
    - Many users strongly miss the V3.2 period and regard V4 as a regression for role-play
  443. 443.
    and creative writing.
  444. 444.
     
  445. 445.
    Suggested directions from users:
  446. 446.
    - long-term memory
  447. 447.
    - persona migration across windows or sessions
  448. 448.
    - official preset-format guidance
  449. 449.
    - a dedicated role-play mode

Yandere Thread - Yandere Applejack (completed)

by Guest

Bonding with Nature

by Guest

The Long and Short of It (RGRE)

by Guest

incest relationships piece of the whole pie (lewd) by Frostybox[...]

by Guest

incest thread piece of the (non-canon) pie, limestone's pie by[...]

by Guest