# CueKFS — paper writing and rebuttal archive This archive holds everything needed to **write the CueKFS paper and defend it during rebuttal**. It is a working archive, not a software release. The runnable code lives in the companion package `author`. --- ## 1. What the project is, for a reader who has never seen it **The task.** A long video can be hundreds or thousands of frames, but a vision-language model can only look at a handful. So a *selector* is given the video and a question, and must return `k` frames (here `k = 8`, `16` or `32`). A separate model, the **reader**, then answers the question using only those frames. The selector never answers anything itself, and it is judged indirectly: by how accurate the reader becomes. This task is called **keyframe selection (KFS)**. **What everyone else does.** Score every candidate frame by how similar it is to the question, keep the top ones. **Why that breaks.** A frame can be relevant without being *literally similar* to the question. Two concrete failures, both measured in Figure 1: 1. The question-level similarity signal is often empty. On the running example, the best-matching frame in a 102-frame video scores `cos = 0.131`, below SigLIP2's own match boundary of `0.147`. Across Video-MME the raw question matches a median of `0.4%` of frames, and for `45%` of questions it matches none at all. 2. Decomposing the question into parts fixes the first problem but creates a second one: a part that is fixed before any evidence is seen can retrieve the wrong thing confidently, and no published selector has a path by which that evidence can get back to the part that produced it. **What CueKFS does.** It is training-free and reframes the problem as matching frames against a set of **dynamic visual cues**. Three steps: | Step | Name | What happens | |---|---|---| | 1 | Decompose | A compact overview of the video is shown to a reasoning VLM, which splits the question into *cues* — short self-contained phrases, each describing one concrete visual element a single frame could show. | | 2 | Cue-Wave-Guided Refinement | Every cue is scored against every frame, tracing a *wave* over time. Peaks and plateaus on that wave are the cue's own evidence. A reasoning VLM sees each cue next to its evidence and may keep, rewrite, split or drop it. Revised cues probe the video again. **This is the core contribution.** | | 3 | Cue Budget Allocation | Cues are ranked by importance, shared evidence is de-duplicated, and the `k` frames are allocated across the surviving cues. Leftover slots are backfilled by farthest-point sampling. | **How well it works.** State of the art on three benchmarks (Video-MME, LongVideoBench, MLVU), three readers (gpt-5.5, Qwen2.5-VL-7B, InternVL3-8B) and three budgets, at a median of two reasoning-VLM calls per question. Headline: Video-MME with gpt-5.5, budget-average **82.70** against WFS-SB's 78.64 and uniform sampling's 77.41. Every term above is defined again, more carefully, in §1 of `01_paper/SUBMISSION_MATERIALS.md`. --- ## 2. What is in this archive ``` 00_source_of_truth/ where the narrative comes from 01_paper/ the manuscript, its source materials, and its figures 02_rebuttal/ evidence for questions reviewers are likely to ask 03_figure_toolkit/ the parts and scripts used to draw figures ``` ### 00_source_of_truth — read this first | File | What it is | |---|---| | `KFS_Final_Presentation.pdf` | The project presentation. **This defines the narrative.** CueKFS is the first half; the second half is a different project (HeroFrame-Bench) and is out of scope here. Note the slides unfold the idea gradually, for an audience; the paper deliberately does not. | | `prompt_that_produced_the_materials_doc.txt` | The instruction that generated `01_paper/SUBMISSION_MATERIALS.md`, kept so it is clear what that document was asked to be. | | `project_one_pager.md` | A short project summary. | ### 01_paper | Path | What it is | |---|---| | `SUBMISSION_MATERIALS.md` | **The main working document.** Self-contained: glossary, abstract plan, introduction, related work, method, results, statistics, figure specifications, provenance for every number, limitations, to-do list. Written so someone with no prior context can read it top to bottom. | | `PAPER_GUIDE.md` | Progress tracking and the writing plan. | | `POSITIONING_AGENTIC_QA.md` | How this work is positioned against agentic video-QA systems. | | `manuscript/` | The LaTeX source (`cuekfs_iclr2027.tex`), bibliography, ICLR style files, build script, and the current compiled PDF. Build droppings (`.aux`, `.log`, `.fls`, `.synctex.gz`) were removed. | | `figures/fig1..fig8/` | One directory per figure. See below. | | `figures/_build/` | The scripts that render the figures. | | `results_tables/` | Where every number in the paper comes from. See below. | **Figures.** Each `figN_*/` directory contains three things: - `full.png` — the complete composed render. This is a **reference for what the figure should say**, not the artwork to paste into the paper. - `full.html` — the same render as editable HTML. - `parts/*.svg` — the individual components, to be rearranged and restyled by hand. | Figure | Parts | |---|---| | `fig1_motivation` | `a_question_wave`, `b_cue_waves`, `c_frozen_vs_loop` | | `fig2_pipeline` | `a_input`, `b_stage1_decompose`, `c_stage2_refine`, `d_stage3_allocate` | | `fig3_step1_decompose` | `a_segmentation`, `b_cue_generation` | | `fig4_step2_cwgr` | `a_wave_anatomy`, `b_batched_review`, `c_reprobe` | | `fig5_step3_allocate` | `a_rank_dedup_allocate`, `b_final_frames` | | `fig6_main_results` | `a_budget_average`, `b_margin_by_budget`, `c_k32_competitors` | | `fig7_ablation` | `a_additive_ladder`, `b_cost_frontier` | | `fig8_behavior` | `a_action_rates`, `b_effect_per_action`, `c_coverage`, `d_uncovered_control` | **results_tables/** — the provenance layer. When the paper states a number, this is where it was computed: | File | What it settles | |---|---| | `downstream_reader_results.md` | **The authoritative main table.** All three benchmarks × three readers × three budgets, for CueKFS, WFS-SB and uniform. Also records the denominator audit, which numbers changed and why, and which cells were deliberately left alone. Read the notes, not just the table. | | `ablation_results_v4.md` | The full ablation programme. | | `abl_tables456.md` | The deeper ablation and analysis tables. | | `experiments_ledger.md` | Running ledger of what was run when. | | `split_analysis_videomme.md` | Accuracy split by where the evidence sits in the video (single moment / multiple moments / whole video) and by whether the retrieval text has a concrete anchor. The `where` axis separates the methods; the anchor axis does not. | | `toolmerge_comparison.md`, `rawmax_results.md`, `h8_research.md` | Side comparisons and probes. | | `analysis_summary.md`, `gpt55_taxonomy_summary.md`, `v46_audit_report.md` | Analysis summaries and an independent audit. | | `system_design_v3.md`, `system_design_v4.md` | The design record: what was tried, what was rejected, why. Useful when a reviewer asks "did you consider X". | | `reproduce_gpt55.md`, `experiment_commands_v4.md` | The commands actually used. | ### 02_rebuttal Has its own README with a reviewer-question-to-file map. See `02_rebuttal/README.md`. ### 03_figure_toolkit The drawing kit: 42 reusable component PNGs in `elements/`, the related-work figure set in `related_work_figs/`, the Python builders in `build_scripts/`, and the rendered slide images in `rendered_slides/`. The `node_modules` directory of the original toolkit was dropped; reinstall with `npm install` if the HTML renderers are needed. --- ## 3. Two things to be careful about **The authoritative numbers are in `01_paper/results_tables/downstream_reader_results.md`, not in the raw run records.** The records under `02_rebuttal/per_question_eval/` are what is physically on disk, and they cover the three benchmarks unevenly: - **Video-MME is complete and exact.** `videomme_jl_j7_seconds1p0_siglip0p98_k{8,16,32}` is the paper's method and reproduces its numbers to the decimal (80.70 / 83.33 / 84.18 over 2699 questions). The reasoning VLM was gpt-5.5. Note that this run reused frozen cues from earlier in the hyper-parameter ladder and re-executed only step 3, which is what J7 tunes; the full provenance chain is documented in the companion package under `release_artifacts/README.md`. - **LongVideoBench and MLVU are not the paper's runs.** What is on disk is `abl_d2_batchreview_prop`, a configuration that predates the J7 search. All 18 of its reader-by-budget cells land 0.3 to 3.5 points below the published numbers, and LongVideoBench additionally predates the fps-densify rerun. Those final runs were executed on separate compute and were not recoverable when this archive was built. Also note that `mlvu × gpt5_5 × k16` on disk is a truncated file holding only two questions; ignore it. So: analyses that need per-question ground truth are safe on Video-MME and should not be run on the LongVideoBench or MLVU records without accounting for the above. **Codenames.** The codebase uses internal names that never appear in paper prose. `j7` is the final configuration; `d2` is the batched-review plus proportional-allocation variant; `propsel` is proportional selection. The mapping is in §10.4 of `01_paper/SUBMISSION_MATERIALS.md`. --- ## 4. Companion package `author` holds the cleaned, runnable code: the method implementation, the configuration files, the baseline converters, and a step-by-step README. Nothing is duplicated between the two archives — code lives only there, figures and rebuttal evidence live only here.