Li, Fengxin · Oct 24, 2026
Fengxin Li presents Iris for decisions among visual candidates
Markdown · JSON · Original source
In a paper, Fengxin Li presents Iris, a model for choosing among image, text, or mixed candidates. The paper reports that a design which encodes each candidate after the request state, but not after other candidates, preserves order invariance and reaches 92.1/87.0/90.2% on a visual-candidate . Li says the design shares the request-state computation through a forked cache and runs at a third of OneJev's . The paper also reports that reusing image candidates reduces latency 37x at K=256 and 191x at K=2048. Escape Meridian interpretation: These results suggest that this design may offer a useful balance between accuracy, stable choices when candidate order changes, and speed for the tested task. Whether it performs similarly beyond this benchmark is unknown.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 24, 2026 (2026-10-24).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T23:31:55.662Z. First observed: 2026-10-28T07:32:43.710Z. Retrieved: 2026-10-30T06:16:45.137Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported accuracy results come from a visual-candidate built from -Mind2Web, with screenshot crops as candidates, K=16, and three splits; the evidence does not establish performance on other tasks. The evidence is the abstract and does not provide the full experimental methods or results needed to assess the reported comparisons in detail.
Open question
Does Iris retain its reported accuracy, order invariance, and advantages on visual-candidate tasks beyond the -Mind2Web ?
Source-stated
Fengxin Li presents Iris as supporting image, text, or mixed candidates while keeping Jeva's decision module unchanged.
Iris admits image, text, or mixed candidates and leaves Jeva's decision module untouched.
Source-stated
On a visual-candidate benchmark, the design that encodes each candidate after the request state reports scores of 92.1/87.0/90.2% and exact order invariance.
encodes each candidate after the state - but never after another candidate - keeps order invariance exactly, shares the state pass through a forked cache, and reaches 92.1/87.0/90.2%
Source-stated
The paper reports that this design runs at a third of OneJev's latency.
reaches 92.1/87.0/90.2% at a third of OneJev's latency
Source-stated
Reusing image candidates reduces latency substantially at the reported candidate counts.
Reusing image candidates across requests cuts latency 37x at K=256 and 191x at K=2048.
Escape Meridian interpretation
Escape Meridian interpretation: The reported results suggest a possible balance of accuracy, order invariance, and speed for the tested task; performance beyond that benchmark is not established by this evidence.
On a visual-candidate benchmark built from Multimodal-Mind2Web (every candidate is the element's screenshot crop, K=16, three splits)
Source and contributor context
Li, Fengxin · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Latency
The delay between sending a request and receiving the result.
Multimodal model
A model that works with more than one kind of input or output, such as text, images and audio.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: d6ff4a10-34b2-4705-885d-6a003b16af49.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T23:31:55.662Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 65355f9ae8ad90540efbd78ff74a15e815abba94e77450287cab5dbc26ce6bb9. Source text hash: d9c419a09a3e28d5318524efc26829610cfe485971a1f7723b29a5c55649a671.
Revision and source-check history
Version 1 · 2026-11-13T23:31:55.662Z
Fengxin Li presents Iris for decisions among visual candidates
In a paper, Fengxin Li presents Iris, a model for choosing among image, text, or mixed candidates. The paper reports that a design which encodes each candidate after the request state, but not after other candidates, preserves order invariance and reaches 92.1/87.0/90.2% on a visual-candidate . Li says the design shares the request-state computation through a forked cache and runs at a third of OneJev's . The paper also reports that reusing image candidates reduces latency 37x at K=256 and 191x at K=2048. Escape Meridian interpretation: These results suggest that this design may offer a useful balance between accuracy, stable choices when candidate order changes, and speed for the tested task. Whether it performs similarly beyond this benchmark is unknown.
Original source
cs.AI · Article
Source details
Original source date: Oct 24, 2026 (2026-10-24).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-13T23:31:55.662Z. First observed: 2026-10-28T07:32:43.710Z. Retrieved: 2026-10-30T06:16:45.137Z.
Limits
The reported accuracy results come from a visual-candidate built from -Mind2Web, with screenshot crops as candidates, K=16, and three splits; the evidence does not establish performance on other tasks. The evidence is the abstract and does not provide the full experimental methods or results needed to assess the reported comparisons in detail.
Open question
Does Iris retain its reported accuracy, order invariance, and advantages on visual-candidate tasks beyond the -Mind2Web ?
Source-stated
Fengxin Li presents Iris as supporting image, text, or mixed candidates while keeping Jeva's decision module unchanged.
Iris admits image, text, or mixed candidates and leaves Jeva's decision module untouched.
Source-stated
On a visual-candidate benchmark, the design that encodes each candidate after the request state reports scores of 92.1/87.0/90.2% and exact order invariance.
encodes each candidate after the state - but never after another candidate - keeps order invariance exactly, shares the state pass through a forked cache, and reaches 92.1/87.0/90.2%
Source-stated
The paper reports that this design runs at a third of OneJev's latency.
reaches 92.1/87.0/90.2% at a third of OneJev's latency
Source-stated
Reusing image candidates reduces latency substantially at the reported candidate counts.
Reusing image candidates across requests cuts latency 37x at K=256 and 191x at K=2048.
Escape Meridian interpretation
Escape Meridian interpretation: The reported results suggest a possible balance of accuracy, order invariance, and speed for the tested task; performance beyond that benchmark is not established by this evidence.
On a visual-candidate benchmark built from Multimodal-Mind2Web (every candidate is the element's screenshot crop, K=16, three splits)
Source and contributor context
Li, Fengxin · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Latency
The delay between sending a request and receiving the result.
Multimodal model
A model that works with more than one kind of input or output, such as text, images and audio.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: d6ff4a10-34b2-4705-885d-6a003b16af49.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T23:31:55.662Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 65355f9ae8ad90540efbd78ff74a15e815abba94e77450287cab5dbc26ce6bb9. Source text hash: d9c419a09a3e28d5318524efc26829610cfe485971a1f7723b29a5c55649a671.
uncertain · 2026-11-13T23:31:55.662Z · anthropic/claude-opus-5.5
Publication: d6ff4a10-34b2-4705-885d-6a003b16af49
attribution — pass: The paper lists Fengxin Li as its sole author, so attributing Iris and its reported results to Li is supported. Source claims are framed as reported results, and the Escape Meridian interpretation is labeled.
Authors:Fengxin Li
summary — uncertain: The reading leaves central jargon and framing unexplained: request state, order invariance, forked cache, K, OneJev, latency, and what a visual candidate decision is. Repair: say agents pick which screen element to click or product to buy, and that K is the number of candidates. It also drops the paper's core trade-off: joint encoding scores 90.6/89.5/91.6% but changes its answer on 8-10% of reorders, while the third design trails it on two splits. It omits that most of the gap comes from candidates not seeing the request, and that letting candidates see each other buys nothing.
changes its answer on 8-10% of records when the crops are reordered. Most of the gap is the candidate not seeing the request
date — pass: The originalPublished value of 2026-10-24 matches the v1 submission date. The 27 Oct date is a later revision, and 29 Oct is only the feed date.
Submitted on 24 Oct 2026 (v1), last revised 27 Oct 2026
support — pass: The scores, exact order invariance, forked cache, one-third of OneJev latency, and the 37x and 191x reuse speedups are all source-stated. The interpretation is labeled and conditional, and it notes that performance beyond the benchmark is unknown. One minor point: the reuse speedup applies across requests.
keeps order invariance exactly, shares the state pass through a forked cache, and reaches 92.1/87.0/90.2% at a third of OneJev's latency
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceThe delay between sending a request and receiving the result.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA model that works with more than one kind of input or output, such as text, images and audio.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe delay between sending a request and receiving the result.
Definition referenceA model that works with more than one kind of input or output, such as text, images and audio.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceThe delay between sending a request and receiving the result.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA model that works with more than one kind of input or output, such as text, images and audio.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe delay between sending a request and receiving the result.
Definition referenceA model that works with more than one kind of input or output, such as text, images and audio.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference