[KartikP]

Lets a BrainScoreModel take part in lab-style studies. Paired with KartikP/brainscore-unified#56, which adds the matching wrappers, tools, registrations and the flagship notebook. The two are pinned to each other (umi-v2.yml here, install/peer-revisions.json there).

What changes

Multi-image trials. TaskContext.present accepts named parts, so one trial can show a sequence before a single answer:

present=('image:instruction', 'image:word_1', 'image:word_2')

• Prompts carry each part's name.
• A multi-image trial is recorded as (presentation, part, neuroid), labelled by part.

Recording on every task path. Each case below used to raise before any compute; it now records the prompt pass:

• during likelihood scoring (the likelihood part must run the prompt on its own first, which is now documented);
• next to a reply with implied probabilities, which are scored first from the same prompt;
• for composite regions;
• in streaming sessions, where each trial's outputs are sent as it ends;
• for extractors without record(), through forward hooks on the module passed as model=.

Memory across trials (contract change).

• Subject.traits is empty by default. A benchmark that needs one lists it in required_traits, and pre-flight marks a subject without it N/A. This is rule 5.10.
• Rule 9.3 reason: input matching allows partial matches, so a memory requirement written as an input would let a feedforward model run a memory task, answering each trial alone. Memory isn't an output either. A declared trait makes such a subject N/A without bending either rule, and existing subjects and benchmarks are unchanged.
• BrainScoreModel(memory='session') declares memory:session and passes earlier trials to reply and likelihood as prompt.history. reset() and the start of a new session clear it.

Fused extractors. extractor={('vision', 'text'): fused} declares each input type and is called once for a stimulus that brings several of them. Single-type extractors still take single-type stimuli.

Implied reaction times. Choice(labels, reaction_time='evidence_margin' | 'surprisal') returns an implied RT on behavior:rt, computed from the readout's or likelihood's label probabilities and labelled with its link. They're scored by correlation, as Futrell2018 scores surprisal and Kar2019 scores solution times. Readout and likelihood parts declare behavior:rt.

Two tests that pin out_channels now include behavior:rt. Two tests that used vision as an example of a family without addresses now use text.

Verification

• New unit tests:
  • test_session_memory.py: a 1-back task that only a model with memory solves; N/A in pre-flight; reset.
  • test_recording_everywhere.py: every path above.
  • test_fused_extractor.py.
  • test_task_responses.py: named parts, implied RTs.
  • test_visual_forms.py: recording per image.
• Full core suite: 1189 passed. The 8 failures fail the same way on unchanged v2 in my environment: a space in the checkout path, and a database test.
• Real weights: Qwen2.5-VL-7B on an A10G, using unified#56.

Check	Result
Fused image and sentence, one pass	parts image, text, final; shape (2, 3, 3584)
1-back on word images	0.44 without memory, 0.78 with memory='session'
Recording during likelihood	(4, 3584), labelled as the prompt pass
Weight edit, then undo	restored bit for bit
Edit to all heads of a layer, then reset()	answers identical to before
Flagship study	see unified#56