gavinlaw/multimodal-understanding-funqa-support-outputs

Finite-hypothesis support-change diagnostic on FunQA

Generated hypotheses, binary scores and measurements from two fixed 4B vision-language models on 71 source-grouped FunQA validation videos, eight uniform windows each. This is a diagnostic adapted from belief-update ideas, not a reproduction of SPIKE-RL, a trained method, or independent test evaluation. All 15,904 calls include hypothesis generation and every baseline/control. See summary.json for the frozen gate result, uncertainty and limitations; negative outcomes are retained.

Source: https://github.com/Nicous20/FunQA directly links https://huggingface.co/datasets/fesvhtr/FunQA . The immutable dataset revision and original CC BY-NC-SA 4.0 license are recorded in manifest.json. We do not redistribute videos, extracted frames, source questions or original annotation intervals. Our generated outputs and code are CC BY 4.0 to the extent we hold rights; this does not relicense original data or models. Derived diagnostic scores require the source annotations for verification.

Selection uses 24 source groups per category, SHA256 ordering, one clip per group. One invalid C interval exceeds the released clip and was excluded before inference without replacement, leaving 24 H / 23 C / 24 M. H/M localization is in frames; C is in seconds. Filenames group original sources but do not certify near-duplicate or pretraining exclusion. Each clip uses 33 uniformly spaced decoded indices and eight four-past/one-current windows, not annotation-selected frames. Fifty-three videos contain both positive and negative windows and support per-video AUC; all 71 contribute peak-hit scores.

Each window generates three hypotheses (temperature 0.8, top_p 0.9, cap 48 tokens). The same model scores each hypothesis from past frames only, from past plus current, and from past plus a repeated final past frame. Two additional queries directly score unexpectedness under current/repeat inputs. These are 3 generation + 11 binary-score calls. Yes/no conditional probability is an uncalibrated model support score; raw vocabulary probabilities and binary mass are retained. Exact prompts, seeds, model revisions and processors are fixed in source and config files.

The main score is mean(support_repeat) minus mean(support_current). It is compared with normalized JS, repeated-frame-adjusted JS, direct unexpectedness and repeated-frame-adjusted direct unexpectedness. Each model must exceed all four by at least 0.05 mean per-video AUC with a positive paired source-bootstrap 95% lower bound, median binary token mass at least 0.8, and hypothesis truncation at most 5%. Both models must pass. These are candidate-screening thresholds, not multiplicity-adjusted confirmatory evidence or proof of probability calibration. Normalization invariance is a known mathematical fact, not a new theorem or evidence that common hypothesis rejection occurs naturally.

Cost summaries include all calls required by each method, even when reused in this experiment. Batch seconds divided by batch size are throughput costs, not online request latency; preprocessing is recorded separately. Binary scoring is a prefill forward pass and generates no answer tokens. Complete per-call costs permit independent accounting.

Three descriptive low-level controls were fixed before unblinding: absolute pixel difference from the previous frame, constant-velocity pixel prediction residual, and per-channel histogram total variation, all on bilinearly resized 64x64 RGB. visual-baseline files contain generated features and derived scores, not original annotation intervals. These are neither learned vision-feature baselines nor SURGE/Swift Sampling reproductions, and do not modify the frozen VLM gate.

Reconstruction on a persistent compute host: run prepare_funqa.py --root /external/funqa, then prepare_support_inputs.py at the same root. This fetches the pinned official source and recreates frames and separate labels. Run run_support_probe.py --root /external/funqa/support-evaluation --model vl|intern --shard 0..3 --deadline-unix , with an explicitly granted GPU per worker and caches configured on persistent storage. The original run used eight A6000 GPUs, one worker per GPU. Run analyze_support_probe.py after all shards finish; it validates input/output hashes before producing summary.json. windows.jsonl.gz uses paths relative to external/funqa and contains no images. The recorded original input-file hash includes original absolute paths; individual inference hashes use frame content hashes and remain independent of those paths.

Related work: https://arxiv.org/abs/2509.23433 (SPIKE-RL); https://arxiv.org/abs/2209.01034 (surprise taxonomy); https://arxiv.org/abs/2605.22678 (Swift Sampling). This dataset does not claim novelty over those methods.