WEBVTT

1
00:00:00.320 --> 00:00:06.640
Imagine searching delivery policies to learn whether batteries
can travel by air, among several related documents.

2
00:00:07.380 --> 00:00:14.400
The retriever finds passages about shipping, damaged products,
and air transport, but their relevance differs.

3
00:00:15.260 --> 00:00:21.320
A reranker must identify which supplied passage answers the
question, then improve the candidate ordering.

4
00:00:22.220 --> 00:00:26.660
These fictional documents describe teaching roles rather than
actual benchmark passages.

5
00:00:27.660 --> 00:00:33.400
Generating candidate identifiers adds computation beyond
estimating the relevance of supplied documents.

6
00:00:34.440 --> 00:00:39.200
CompRank directly scores candidates while reducing the tokens
needed for comparison.

7
00:00:39.740 --> 00:00:45.080
CompRank places independent document blocks before a query
decision block that scores the candidates.

8
00:00:45.980 --> 00:00:52.240
Tokens inside each document use causal self-attention, building
a representation from that document's own context.

9
00:00:53.140 --> 00:00:58.400
Cross-document attention is prohibited, so document encoding
does not depend on neighboring candidates.

10
00:00:59.420 --> 00:01:05.060
Local positions avoid document-order offsets, reducing one
source of candidate-permutation sensitivity.

11
00:01:06.000 --> 00:01:11.780
The query block still sees all document blocks, maintaining a
global candidate view when relevance is estimated.

12
00:01:12.560 --> 00:01:18.120
Independent document states can therefore be reused in principle
across queries and reordered candidate lists.

13
00:01:18.700 --> 00:01:25.580
Independent encoding removes cross-document interactions, but
query attention still scales with document tokens.

14
00:01:26.420 --> 00:01:33.500
Compression exposes only selected document key-value states to
the query block, reducing the final interaction cost.

15
00:01:34.680 --> 00:01:41.140
Document representations stay unchanged; selection controls
which states query-to-document attention sees.

16
00:01:42.100 --> 00:01:47.760
Segment-wise selection retains regularly spaced tokens, covering
the document rather than only its end.

17
00:01:48.720 --> 00:01:54.660
The illustrated position numbers show regular spacing, without
asserting which particular words an experiment retained.

18
00:01:55.520 --> 00:02:01.780
The default Step-ten strategy uses that coverage to balance a
smaller attention budget against ranking quality.

19
00:02:02.480 --> 00:02:08.320
Scoring aggregates query attention on retained document tokens,
treating that mass as relevance evidence.

20
00:02:09.200 --> 00:02:15.100
Attention is normalized across the retained document tokens and
then aggregated into a score for each candidate.

21
00:02:15.840 --> 00:02:22.100
The main Mistral configuration uses one decision token,
averaging scores across attention heads at a fixed layer.

22
00:02:23.120 --> 00:02:26.960
Inference directly sorts supplied documents by their
attention-derived scores.

23
00:02:27.440 --> 00:02:32.580
No identifier sequence is generated; scoring avoids sequential
decoding of a ranking string.

24
00:02:33.240 --> 00:02:38.820
Its output is candidate order, not reconstructed document text
or a generated answer to the original query.

25
00:02:39.587 --> 00:02:46.047
Identifier generation uses the full vocabulary, including tokens
unrelated to valid input candidates.

26
00:02:47.067 --> 00:02:52.847
CompRank uses attention-derived scores to form a pointer
distribution over the input candidates.

27
00:02:53.667 --> 00:03:00.027
The CopyNet-style loss supervises the relevant candidate,
training the same scores inference later sorts.

28
00:03:01.207 --> 00:03:07.247
Candidate selection replaces identifier generation, matching
supervision to the reranker decision.

29
00:03:08.247 --> 00:03:14.827
Experiments use Mistral seven B with thirty-document training
lists and positive examples derived from MS MARCO.

30
00:03:15.667 --> 00:03:23.087
BEIR evaluation compares rerankers on the same one hundred BM
twenty-five retrieved candidates across seven datasets.

31
00:03:23.627 --> 00:03:32.827
Table one compares compression strategies within CompRank,
measuring average NDCG at ten over seven BEIR datasets.

32
00:03:33.627 --> 00:03:40.307
The full-token version retains all document tokens and scores
thirty-nine point seven on the reported average metric.

33
00:03:41.207 --> 00:03:47.287
Step-ten retains ten point two percent of tokens and scores
thirty-nine point two, half a point lower.

34
00:03:47.987 --> 00:03:53.867
Token retention measures attention budget; NDCG measures ranking
quality, so they have different meanings.

35
00:03:54.547 --> 00:04:00.787
This averages seven dataset scores, rather than measuring task
completion or promising every-cell improvement.

36
00:04:01.487 --> 00:04:07.447
The matched test supports structured coverage, while stronger
compression still trades away ranking quality.

37
00:04:08.047 --> 00:04:14.267
A TREC-COVID test trains on thirty-document lists, then
evaluates up to five hundred without further fine-tuning.

38
00:04:15.007 --> 00:04:21.027
At five hundred candidates, generation-based Direct-List takes
one hundred fourteen point five seconds per query.

39
00:04:21.947 --> 00:04:27.987
Full-token CompRank takes sixteen point two two seconds;
Step-ten takes twelve for the same candidate count.

40
00:04:29.187 --> 00:04:34.967
These end-to-end timings use one A one hundred GPU with eighty
gigabytes of memory and batch size one.

41
00:04:35.927 --> 00:04:41.307
Measured evaluation recomputes document blocks; optimized
caching remains a separate design opportunity.

42
00:04:42.167 --> 00:04:48.107
The gain combines decoding-free scoring and compressed
attention, rather than unimplemented serving optimizations.

43
00:04:48.647 --> 00:04:55.007
Return to policy search: choose useful supplied evidence, rather
than write a document identifier sequence.

44
00:04:55.947 --> 00:05:02.307
CompRank independently encodes documents, compresses
query-visible states, and ranks attention scores.

45
00:05:03.267 --> 00:05:08.847
Candidate-level supervision trains those scores, aligning the
learning objective with inference-time sorting.

46
00:05:09.867 --> 00:05:15.607
Computation targets sparse comparison signals, instead of dense
interactions between every document pair.

47
00:05:16.587 --> 00:05:22.327
Quality and latency evidence is configuration-specific; caching
potential differs from measured serving.

48
00:05:23.287 --> 00:05:29.607
Align representation, computation, and supervision with the
relevance decision the retrieval system needs.