**An update to the Token Importance Scoring series — new passage reordering capability, transfer learning findings, and open-source checkpoints.** Following the initial [TIS release]( Presenting TIS (Token Importance Scoring) - A new way to compress KV cache ), we’ve significantly expanded TIS capabi…

Hi, I read your reply and I will have a read and get back to you. Thanks.

I think this is much easier to verify now: The public surface is materially easier to inspect than it was earlier. The repository now exposes the scorer paths, evaluation scripts, saved result files, a manifest, compatibility checks, and substantially more checkpoint/training metadata. That makes …

[image] John6666: a globally reliable high-first direction; a globally reliable low-first direction; calibrated query-specific passage relevance; a reproducible TIS-specific ranking advantage over BM25 or TF-IDF; stable generator-facing utility from arithmetic-mean high-first ordering; transfer…

Hi, A fold replaces each group by its mean, and the mean of n samples has standard deviation σ/√n. So initialising a matrix and then projecting leaves the weights smaller by √(tying ratio). Draw the initialisation with σ·√(group size), so the weights carry the intended scale after projection rathe…

VERSION 2.2 IS HERE!! HuggingFace — new repo oldman-dev/tis-v2.2-passage-reranker: tis_components.pt (244 MB, SHA256: d26012b…) — QueryAwareImportanceHead trained with supervised MS-MARCO labels metadata.json — full training config, scoring contract, eval results, checkpoint identity README.md —…

This looks like a much more solid direction overall: I think v2.2 is a much cleaner way to resolve the direction question than trying to infer the semantics of the older ERT scalar after the fact. With the new supervised branch, the objective itself says, in effect, score(relevant passage) > sco…

As usual, thanks for the deep dive! I was preparing a 2.3 version already, but I decided to add your recommendations about the explicit boundary to it … New training sprints are being ran right now (this time with 2000 steps, for a stronger training set) and then the appropriate tests using the foun…

And it’s published! [image] New checkpoints in HuggingFace and a new code version ready at the GitHub link! If no major problems arise when reproducing, this closes TIS2.0 series. Now I will start: TIS3.0 - maybe focused on DRAFTER related problems! TIS+PIN - a parallel distilation where I wil…

Umm… this has gotten a lot better, but I still have this vague feeling that something may or may not be missing here:sweat_smile:: I spent some more time trying to reproduce the v2.3 result from the public artifacts. The short version is: I do not think model quality itself is a reason to hold u…

As always, thanks for the detailed feedback! I will need to stop relying so much in the “semi-automatic” packager and try to make the packaging of the public versions myself or micro-manage them! As soon as I have some time, I will start preparing the correct upload, because the results are correct…

TIS 2.0: Token Importance Scoring Now Eliminates Position Bias in RAG

Research

oldman-dev August 13, 2026, 12:48pm 23

Well! I have uploaded to GitHub the missing parts and documentation! I hope everything is good now, but I liked your MRR results, higher than mine during my testing/training!

Topic		Replies	Views
Presenting TIS (Token Importance Scoring) - A new way to compress KV cache Research	22	719	July 30, 2026
Seeking feedback on token-block context selection for long-sequence QLoRA fine-tuning Research	13	425	September 13, 2026
MosAIc: a reconfigurable surface where an LLM emits a layout, not just text Show and Tell	7	261	July 11, 2026
Need feedback on my research Research	1	118	July 13, 2026
SplatRAG 3+ Visualizer: SciFact 0.8804 / 0.7624 with 5,000 slang rows in the same memory Show and Tell	2	80	August 23, 2026

TIS 2.0: Token Importance Scoring Now Eliminates Position Bias in RAG

Related topics