A video search system can find a clip that sounds right, but still has no idea whether it should trust that answer. This paper gives each proposed clip a separate confidence score, including permission to return nothing.
Video temporal grounding supports applications such as video search, content review, and automated editing by localizing events described in natural language. Yet existing generative models typically output timestamps without explicit intervallevel confidence scores to guide candidate selection. We separate candidate generation from acceptance by scoring individual intervals within the original decoding pass. A lightweight confidence head reads pooled decoder states, providing an explicit score trained for interval selection. Offline verifier scores supervise the head on fixed candidate sequences, and temporal-overlap labels adapt it to current rollouts during reinforcement learning. GT-anchored candidate-pool supervision and set-level optimization train the generator. The resulting scores support ranking, threshold-based selection, and rejection without invoking an external verifier at inference. On a fixed OMTG-Bench candidate pool, confidence raises query-macro Recall@0.5 from 9.95% to 14.42% over generation order at a 10% global return budget, and from 26.48% to 31.12% at a 25% budget. The continuous scores let downstream applications adjust return budgets or acceptance thresholds to match their precision–recall preferences, without regenerating candidate intervals.
Transcript
A video search system can find a clip that sounds right—but still has no idea whether it should trust that answer. This paper gives each proposed clip a separate confidence score, including permission to return nothing. Video temporal grounding supports video search, content review, and automated editing by locating events described in natural language.
But existing generative models typically output timestamps without explicit confidence scores for each interval. The approach separates candidate generation from acceptance by scoring individual intervals during the original decoding pass, giving the acceptance step interval-level scores.
Those scores support ranking, threshold-based selection, and rejection without an external verifier at use time. Applications can adjust return budgets or acceptance thresholds without regenerating candidate intervals. Each generated interval is paired with a continuous confidence score, using information already produced while the response is generated.
The same response can then be judged without generating timestamps again. Applications can adjust an acceptance threshold, rank candidates by reliability, or return an empty result when no candidate passes the threshold.
Candidate generation determines what can be returned; confidence supplies the acceptance rule. The system provides controllable grounding from a single decode. It separates candidate generation from acceptance and scores intervals from the original decoder states.
That makes ranking, threshold-based selection, and rejection possible without decoding again or using an external verifier during use. At use time, the system scores intervals from states saved during timestamp generation, without adding confidence tokens. The scoring reuses features from the same decoding process.
Generating additional candidates still requires the cost of producing their tokens. The single-pass scoring helps choose among candidates already generated; it does not make extra candidates free. Confidence improves recall over simply following generation order by 4.47 points at 10 percent retention and 4.64 points at 25 percent retention.
The benefit is largest under tight budgets: as more of the fixed pool is returned, selection has less room to help. In other words, confidence matters most when only a small share of possible results can be shown. In a rejection comparison, the native system keeps false alarms at rates from 66.04 to 91.17 percent on mismatched queries.
At test-selected thresholds, the system reduces that range to 13.24 to 21.35 percent. The system also improves the combined balance between correct results and false alarms by 1.40 to 5.63 points, although the gains vary by dataset and some settings trade recall for precision.
A useful retrieval system must know when not to answer. Here, it accepts both queries that truly belong to the video, but returns no result for both deliberately mismatched queries—even though it generated candidates for all four. The confidence value is not meant to mean that the same number is a percentage chance of being correct.
Instead, a threshold is fitted to an observed precision goal. The transferred settings keep recall at 55.75 percent and 33.33 percent, so higher precision comes with reduced coverage. The study also leaves open how much changes in threshold choice, group composition, and the score-to-correctness relationship contribute.
The results show that generation and selection are distinct capabilities: the value of a candidate pool depends on how its intervals are prioritized and accepted. Fixed-pool experiments show improved recall under equal return budgets, with the largest gains at tight budgets.
Explicit confidence makes the output policy adjustable after decoding. The same candidates can serve different return budgets and acceptance requirements, so a video search tool can show likely matches, demand stricter evidence, or safely show nothing. The key idea is to separate finding possible moments from deciding which ones deserve to be shown.
That makes the same search results adjustable for careful review, fast editing, or situations where false matches are costly.
A derivative work by Paperi · AI-generated script, voice and captions
· pages and figures unaltered
Made with Paperi.
Drop in a research PDF — get a narrated video walkthrough like this one,
with highlights that follow the narration. Free to start.
A model can score as if it matches the brain even after the brain information has been destroyed. And a language pattern can look real when the test measuring it is failing.A model can score highly against the brain, or transfer beautifully across languages, while the score mostly reflects something other than the structure researchers hoped to measure. This paper tests exactly how much correspondence survives when that structure is removed.
A robot may have a floorplan, yet still need a huge prepared database before it can tell where it is. FreeLoc asks a simpler question: can the robot query the map directly, as it moves?What if a robot could localize indoors using only a floorplan, without building a giant image or ray database first? FreeLoc makes that idea practical by querying geometry online and refining rough guesses with diffusion.
An AI can remember something about you perfectly and still bring it up at exactly the wrong moment. The real challenge is not remembering more, it is knowing when a memory belongs in the conversation.An AI companion can remember something perfectly and still use it badly. This paper asks the harder question: before bringing up a memory, should the system decide whether that memory belongs in this moment at all?
4:48
