Is "Pure" NNUE strong also in losing position?

[RogerC]

It seems that pure NNUE evaluations (with last official net nn-baeb9ef2d183.nnue) finds quickly solutions on bad positions. Some examples here. In those positions, standard SF (with hybrid eval) is not finding the best move to get out of losing positions because the imbalance is so important that it uses only standard evaluation and not NNUE eval any more. If you "force" CFish with "pure" NNUE UCI option to use only NNUE eval, it finds all the best moves in less than 5 seconds.

May be SF developpers should not switch to standard evaluation unless the imbalance between player and opponent is strickly positive and not absolute positive (eval > threshold and not abs(eval) > threshold). This should make SF use more often Pure NNUE eval when in negative imbalance so that it would find quickly a way out of losing positions. A try should be done in SF testing framework.

[d]rn1qrnk1/p4pp1/1p1pp3/6P1/2Pp1PN1/2PQ4/P5P1/2KR3R w - - 0 1 bm Nh6+
[d]4k3/4Pp2/1P1p1P1P/pPpPpK2/pr2pbP1/7r/3RP3/NN5b w - - 0 1 bm Rb2
[d]r1n2N1k/2n2K1p/3pp3/5Pp1/b5R1/8/1PPP4/8 w - - 0 1 bm Ng6

[RogerC]

Another losing position where SF (with hybrid eval) is blind to a way out (bm e6). CFish with "pure" NNUE eval finds it in a second :
[d]rn2kbnr/pp2ppp1/2p5/3pP1Bp/3P3P/3Q4/PqP2PP1/RN2K1NR w KQkq - 0 1 bm e6

[peter]

Another losing position where SF (with hybrid eval) is blind to a way out (bm e6). CFish with "pure" NNUE eval finds it in a second :
[d]rn2kbnr/pp2ppp1/2p5/3pP1Bp/3P3P/3Q4/PqP2PP1/RN2K1NR w KQkq - 0 1 bm e6

Thanks for the positions, knew 1st one from a game Eduard Nemeth (member here) against Shredder, not sure about the year, something between 2008 and 2013, 3rd one as a Zugzwang study of Axel Akerblom, which is used too in Eret test suite as nr.2.

This last one is a position from Caro- Kann, with still the older one moves 8.Ne2(?!) and 8.Nf3(?!) in some books, e6(!) had been played then in a corr. game in 2015 for the first time in corr. that I know of:
[pgn]
[Event "corr BI-2015-P-00007"]
[Site "Lechenicher SchachServer"]
[Date "2015.04.16"]
[Round "?"]
[White "Bogomolov, Alexander"]
[Black "Koehler, Josef"]
[Result "1-0"]
[ECO "B12"]
[WhiteElo "2035"]
[BlackElo "1955"]
[Annotator "Martan,Peter"]
[PlyCount "123"]
[EventDate "2015.04.22"]
[EventType "tourn (corr)"]
[Source "H.Bellmann"]
[SourceDate "2020.10.15"]

1. e4 c6 2. d4 d5 3. e5 Bf5 4. h4 h5 5. Bg5 Qb6 6. Bd3 Bxd3 7. Qxd3 Qxb2 8. e6
(8. Ne2 $6 {Hiarcs} Na6 9. O-O Nb4 10. Qd2 e6 $15) 8... Nh6 9. Nf3 Na6 (9...
Qb5 10. exf7+ $14) 10. O-O Nb4 11. Qd2 a5 12. Nc3 Qxc2 13. Qf4 Qg6 14. Rae1 Rc8
15. exf7+ Nxf7 16. Re2 Nd3 17. Qe3 e5 18. Rd1 Bb4 19. Rxd3 e4 20. Rd1 O-O 21.
Ne5 Qf5 22. Nxf7 Rxf7 23. Qg3 Qd7 24. Rb1 Bd6 25. Qe3 Rcf8 26. Nd1 Rf3 27. Qc1
Rd3 28. Qb2 Bb4 29. Ne3 Kh7 30. Kh1 b6 31. Kg1 Qe6 32. Nc2 Bc3 33. Qc1 Qg4 34.
Qf1 b5 35. Rc1 b4 36. Re3 Qf5 37. Rb1 Kg8 38. Be7 Rf7 39. Bc5 Rd2 40. Re2 Rd3
41. a3 bxa3 42. Bxa3 Qg4 43. Rb8+ Kh7 44. Rb6 Bd2 45. Bd6 Qg6 46. Bg3 a4 47.
Rxd2 Rxd2 48. Qc1 Rd3 49. Ne3 Rb3 50. Ra6 Rfb7 51. Kh2 Rb2 $2 (51... Rb1 {
[%eval 51,41]}) 52. Nxd5 Rb1 53. Qe3 Rd1 54. Nc3 Rd3 55. Qc1 Qh6 56. Qe1 Re7
57. d5 a3 58. dxc6 Qe6 59. Ne2 Qc4 60. Qa5 Rd5 61. Qxa3 Re8 62. Nf4 1-0
[/pgn]
After that I didn't find any more games with the line on Herbert Bellmann's DVD Corr 2020.
In Cerebellum3Merge.bin it's already stored as best move in the position.