NeurIPS / EMNLP / LoG2026

Self-play. Retrieval.
Steering. Evaluation.

Self-play

Learning without
curated answers

Retrieval

Inside text + image
attention heads

Steering

Composing multiple
behaviours

Evaluation

Beyond headline
benchmark scores

Graph learning

Interfaces for relational
generalisation

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

Self-play, retrieval, and evaluation.

NeurIPS

SCOPE

Self-play for open-ended tasks

Cyrus Kwan et al.

LOCOS

What retrieval heads actually write

Aryo Gema et al.

Composed image retrieval

When benchmarks admit shortcuts

Matteo Attimonelli et al.

AI psychophysics

Grounding behavioural evaluation

Jonas Leo Mueller, Dario Zanca et al.

EMNLP

Multimodal retrieval heads

Can retrieval heads see images?

Aaron Branson Cigres Li et al.

Hidden VLM instability

Same answer, different representations

Farooq Ahmad Wani et al.

Composing steering vectors

Controlling behaviours without interference

Giwon Hong et al.

LoG

Graph-to-table interfaces

What controls relational generalisation?

Alessandro Cagiano et al.

Preprints

SPSD

Self-play search distillation

Lorenzo Molfetta et al.

FACE-Eval

Chain-of-thought faithfulness

Aryo Gema et al.

VLM frequency filtering

Why verbose prompts help robustness

Farooq Ahmad Wani et al.

NeurIPS / EMNLP / LoG2026
NeurIPS

SCOPE

Figure from SCOPE

Cyrus Kwan et al.

NeurIPS

LOCOS

Figure from LOCOS

Aryo Gema et al.

NeurIPS

Composed image retrieval

Figure from Composed image retrieval

Matteo Attimonelli et al.

NeurIPS

AI psychophysics

Figure from AI psychophysics

Jonas Leo Mueller, Dario Zanca et al.

EMNLP

Multimodal retrieval heads

Figure from Multimodal retrieval heads

Aaron Branson Cigres Li et al.

EMNLP

Hidden VLM instability

Figure from Hidden VLM instability

Farooq Ahmad Wani et al.

EMNLP

Composing steering vectors

Figure from Composing steering vectors

Giwon Hong et al.

LoG

Graph-to-table interfaces

Figure from Graph-to-table interfaces

Alessandro Cagiano et al.

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

Learn

NeurIPS

SCOPE

Self-play for open-ended tasks

EMNLP

Composing steering vectors

Controlling behaviours without interference

LoG

Graph-to-table interfaces

What controls relational generalisation?

Inspect

NeurIPS

LOCOS

What retrieval heads actually write

EMNLP

Multimodal retrieval heads

Can retrieval heads see images?

Evaluate

NeurIPS

Composed image retrieval

When benchmarks admit shortcuts

NeurIPS

AI psychophysics

Grounding behavioural evaluation

EMNLP

Hidden VLM instability

Same answer, different representations

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026
ML

Learning.
Retrieval.
Control.

SCOPE

Cyrus Kwan et al.

NeurIPS

LOCOS

Aryo Gema et al.

NeurIPS

Composed image retrieval

Matteo Attimonelli et al.

NeurIPS

AI psychophysics

Jonas Leo Mueller, Dario Zanca et al.

NeurIPS

Multimodal retrieval heads

Aaron Branson Cigres Li et al.

EMNLP

Hidden VLM instability

Farooq Ahmad Wani et al.

EMNLP

Composing steering vectors

Giwon Hong et al.

EMNLP

Graph-to-table interfaces

Alessandro Cagiano et al.

LoG
PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

Learning from self-play.
Looking inside models.
Questioning the tests.

SCOPE self-play method figure

SCOPE / Cyrus Kwan et al.

NeurIPS

SCOPE
LOCOS
Composed image retrieval
AI psychophysics

EMNLP

Multimodal retrieval heads
Hidden VLM instability
Composing steering vectors

LoG

Graph-to-table interfaces

PreprintsSPSD / FACE-Eval / VLM frequency filtering

NeurIPS / EMNLP / LoG 2026

How models learn.
What they retrieve.
How we evaluate.

Preprints

SPSD / FACE-Eval
VLM frequency filtering

NeurIPS

SCOPE
LOCOS
Composed image retrieval
AI psychophysics

EMNLP

Multimodal retrieval heads
Hidden VLM instability
Composing steering vectors

LoG

Graph-to-table interfaces

NeurIPS / EMNLP / LoG2026

NeurIPS

SCOPE

Cyrus Kwan et al.

LOCOS

Aryo Gema et al.

Composed image retrieval

Matteo Attimonelli et al.

AI psychophysics

Jonas Leo Mueller, Dario Zanca et al.

EMNLP

Multimodal retrieval heads

Aaron Branson Cigres Li et al.

Hidden VLM instability

Farooq Ahmad Wani et al.

Composing steering vectors

Giwon Hong et al.

LoG

Graph-to-table interfaces

Alessandro Cagiano et al.

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

Self-play. Retrieval. Evaluation.

NeurIPS

SCOPE

LOCOS

Composed image retrieval

AI psychophysics

EMNLP

Multimodal retrieval heads

Hidden VLM instability

Composing steering vectors

LoG

Graph-to-table interfaces

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

Learning and model behaviour.

NeurIPS

SCOPE

Self-play for open-ended tasks

Cyrus Kwan et al.

NeurIPS

LOCOS

What retrieval heads actually write

Aryo Gema et al.

NeurIPS

Composed image retrieval

When benchmarks admit shortcuts

Matteo Attimonelli et al.

NeurIPS

AI psychophysics

Grounding behavioural evaluation

Jonas Leo Mueller, Dario Zanca et al.

EMNLP

Multimodal retrieval heads

Can retrieval heads see images?

Aaron Branson Cigres Li et al.

EMNLP

Hidden VLM instability

Same answer, different representations

Farooq Ahmad Wani et al.

EMNLP

Composing steering vectors

Controlling behaviours without interference

Giwon Hong et al.

LoG

Graph-to-table interfaces

What controls relational generalisation?

Alessandro Cagiano et al.

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

SCOPE

Cyrus Kwan et al.

NeurIPS

LOCOS

Aryo Gema et al.

NeurIPS

Composed image retrieval

Matteo Attimonelli et al.

NeurIPS

AI psychophysics

Jonas Leo Mueller, Dario Zanca et al.

NeurIPS

Multimodal retrieval heads

Aaron Branson Cigres Li et al.

EMNLP

Hidden VLM instability

Farooq Ahmad Wani et al.

EMNLP

Composing steering vectors

Giwon Hong et al.

EMNLP

Graph-to-table interfaces

Alessandro Cagiano et al.

LoG
PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026
MLreasoning / retrieval
evaluation

NeurIPS

SCOPE
LOCOS
Composed image retrieval
AI psychophysics

EMNLP

Multimodal retrieval heads
Hidden VLM instability
Composing steering vectors

LoG

Graph-to-table interfaces

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

Self-play, retrieval, and steering.

NeurIPS

SCOPE
LOCOS
Composed image retrieval
AI psychophysics

EMNLP

Multimodal retrieval heads
Hidden VLM instability
Composing steering vectors

LoG

Graph-to-table interfaces

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

Learn / control

Manuscript figure illustrating Learn / control

SCOPE

NeurIPS

Composing steering vectors

EMNLP

Graph-to-table interfaces

LoG

Inspect

Manuscript figure illustrating Inspect

LOCOS

NeurIPS

Multimodal retrieval heads

EMNLP

Evaluate

Manuscript figure illustrating Evaluate

Composed image retrieval

NeurIPS

AI psychophysics

NeurIPS

Hidden VLM instability

EMNLP
PreprintsSPSD / FACE-Eval / VLM frequency filtering
2026

{NeurIPS, EMNLP, LoG}

Self-play, retrieval, steering, evaluation,
and relational generalisation.

PreprintsSPSD / FACE-Eval / VLM frequency filtering
NeurIPS / EMNLP / LoG2026

NeurIPS

SCOPE

LOCOS

Composed image retrieval

AI psychophysics

Composed image retrieval examples

EMNLP

Multimodal retrieval heads

Hidden VLM instability

Composing steering vectors

Multimodal retrieval-head method

LoG

Graph-to-table interfaces

Graph-to-table method
PreprintsSPSD / FACE-Eval / VLM frequency filtering