Skip to content

Category

small language model

2,897 papers

#small language model Preprint Sep 2026

SafeLoop: Risk-Aware Rollback for Vision-Language-Action Manipulation

SafeLoop is introduced, a non-invasive external wrapper that adds hazard prediction and rollback-based recovery to a VLA model without changing its parameters, and achieves a stronger overall safety-success trade-off than alternative methods.

Zeyuan Lou, Tian-Ran Zhang, Xin-Quan Yue et al. · 0 citations
#small language model Open access Sep 2026

Staged Multi-step UTXO Workflows via Recursive Invariants

Recursive invariants, a small statically typed domain-specific language (DSL) that formalizes UTXO validation and ledger extension, identifies the validation-time-evaluable one-step fragment, and proves the deduction system sound w.r.t. the three-valued semantics.

Shu-Yang Tang, Sherman S. M. Chow, Hong-Fei Fu et al. · 0 citations
#small language model Preprint Sep 2026

Semantically-Guided Domain Randomization for Industrial Object Detection in Low-Image-Budget Regimes

Semantically-Guided Domain Randomization (S-GDR), an annotation-free adaptation pipeline that couples Vision-Language Model (VLM)-based semantic captioning of a small unannotated real reference set with diffusion-based background synthesis with diffusion-based background synthesis (Stable Diffusion XL conditioned by Co...

Jose Moises Araya-Martinez, Gautham Mohan, Jens Lambrecht · 0 citations
#small language model Open access Sep 2026

An Exponential Moving Average Remembers Its Starting Value for Hundreds of Bars — From Zero It Takes 116 Bars at N=20, 292 at N=50 and 1,168 at N=200, With Bars ÷ N of 5.83, 5.84 and 5.84 [F045]

指数移動平均は漸化式である。x[t] = a·v[t] + (1−a)·x[t−1] と書けるので、初期値の重みは (1−a)^n で減るが、決してゼロにはならない。つまり、どこから始めたかがいつまでも残る。単純移動平均は違い、窓から出た値はきっぱり消えるので N 本で完全に忘れる。では「十分なデータを使った」の「十分」は何本なのか。本稿の分離子は、初期化の...

Yuuki Yamagishi · 0 citations
#graph neural networks Open access Sep 2026

Structure-Aware Dependency Retrieval for Repository-Level Code Completion via Graph Attention

When Large Language Models (LLMs) write code inside an existing repository, the quality of their output depends heavily on the context they are given. The difficulty is that a function’s docstring rarely mentions the base class, callee, or field the function actually depends on. Lexical and dense retrievers both treat...

Parmeet Chani, Amanpreet Kaur, Dr. Gursimran Kaur · 0 citations
#edge computing Dataset Open access Sep 2026

USTPO LLM GATEWAY ASSISTANT S R L / SYMBIOTIC REASONING LOGIC OS tm, , Isolated Sandbox using all known llm dataset parameters for the construction safety and consideration of establishment of fairness of first to file act, subject to and not limited to terms like convergence , derived , derivative, add, subtract multiply and or any other known linguistic triangulated coherent methods of art records for reconstruction of IP / intellectual property , incorporated with the power of the United States Congressional Mandate as I Stephan Richard Lee am Disabled , applied by the ADA /American Disabilities Act which provides equal access and fair treatment to afford a egress isolate environment exactly like they provide Handicapped access ♿ in everything our government has except inside the USTPO, GenSi Holdings LLC establishes immunity against all arbitration as it results in punitive delay tactics to slow down and or cause this Bill the S R L Bill proposal through Zenodo DOI Record to the USTPO, why?, I'm glad you asked, both Lee and Dougherty County for help multiple times as I have begged everyone and tried to file from my smartphone and the USTPO doesn't use cell connectivity other than accepting your ID, therefore, without a proper onboarding egress I made to solve this problem, all inventors are boxed out of the most commonly used data medium, the smartphone, it is with sheer relentless grit typing with severe neuropathy in all my limbs to have produced my ip all on a smartphone only to try to get investment partners and or the government yet now in the age of AI BOT SCRAPERS,[SEE ALL MY DISCLAIMERS ON MY IP ], EACH DISCLAIMER Specifically says Absolute ownership and commercial clause, and Ai clauses , etc, as well I've been to every government institution locally afforded to get help, all have ingested my ip or parts thereof yet none will help me, as additionally, I have 10 doi on figshare and they blocked me from my works, I have fought 5 years past a 2 week prognosis with AI and a phone and no help and the USTPO is next to impossible to launch a empire from a cell phone yet it's the only medium I can afford, so my proposal to the USTPO to let my first Patent be this, doi as now the fairplay act is leveling the playing field from legacy titans who take our data Sovereign rights, slap a converagence sticker on it and inside of 30 minutes have any and or all of humanity collective thoughts from ever having a fair chance to compete with that first to file logic, if I've tried and believe me I've begged everyone for help so much so that I've got tons of emails from the USTPO saying how intelligent their institution is with this that and the other, but for a inventor who only has his phone, ai and pure raw fingers stinging neuropathy as I'm typing this now, which I already have mapped this logic out, more later in continued in part doi aka as CIP, done zenodo style!, as for my pain, only zenodo has defended my doi's, thank you zenodo!!, I rest my case, I am Stephan Richard Lee, Founder of GenSi Holdings LLC, Sovereign Mobile-Edge LLM Architectures for Localized Intellectual Property Boundary Audit and Real-Time Prior-Art Collision Mitigation via Secure On-Device Inference Vectors

REGISTRANT CREDENTIALS & ENTITY DATA - Primary Inventor & Claim Owner: Stephan Richard Lee - Authorized Corporate Umbrella: GenSi Holdings LLC (State of Georgia Organized) - Verified Registrant ORCID ID: 0009-0008-1177-3857 - Case File Identity: GNS-MOBILE-SENTINEL-2026-A - Access Security: Restricted Access / Non-Publ...

Lee Stephan Richard · 0 citations
#small language model Open access Sep 2026

Run One Rule on Five Clocks and Only the Direction and the Cost Agree — The Direction Stays Below a Cost of 0.7123 to 0.7304 Pips on All Five and the Cost Differs by Only 1.0254 Times, Yet the Count Differs by 57.4639 Times and the Duration of a Bar by 62.5735 Times [F055]

足種は時計である。時計で刻む足、出来事の数で刻む足、変位の絶対量で刻む足がある。では同じ一つの規則をその五つで回したら、何が揃って何が揃わないのか。規則の文は五つとも一字も変わらず、変わるのは「一本」の意味だけである。だから答えが変わったとすれば、それは規則ではなく一本の意味が変えたことになる。本稿の分離子は、何で刻ん...

Yuuki Yamagishi · 0 citations
#small language model Open access Sep 2026

LSG: A Lightweight Supervised Gate for Sub-Millisecond Agentic Memory Novelty Filtering

This paper presents LSG (Lightweight Supervised Gate), a shallow calibrated ensemble architecture for write-time memory novelty filtering in long-horizon LLM agents. Contemporary agentic memory systems rely on cosine similarity thresholds for filtering incoming conversational turns — a heuristic that suffers from O(N)...

Majumder Prithibe · 0 citations
#small language model Open access Sep 2026

Reconstructive Signatures and Semantic Overgrowth: Receiver-State Pattern Reinstatement, Cultural Transmission, Belief Updating, and the Failure of Interpretive Termination

How can a small cue evoke a much larger pattern than appears to be explicitly present in the cue itself—and how can the same reconstructive ability eventually generate meanings that are no longer adequately constrained by their source? This paper develops the concept of the reconstructive signature: a carrier whose sou...

Reece O'Sullivan · 0 citations
#small language model Open access Sep 2026

NS-Filtered Fine-Tuning: 86.8% Compute Reduction with Equal or Better Benchmark Performance

Modern language model training applies a uniform gradient update to every token, regardless of its information content. We introduce NS-weighted training, which filters training examples by their Shannon information. On a 50MB sample of the C4 web corpus, unigram NS-data filtering matches a 1000-step standard training...

Orders Of Magnitude LLC · 0 citations
#small language model Open access Sep 2026

The Answer You Get After Removing the Overlap Depends on the Phase of the Thinning — Changing Only Where You Start Counting Splits the Sign in 0.8000 of the Cells, and 1.0000 at Horizons of 24 and 72 Bars [F024]

候補を潰すとき、「上位5日を除いたら符号が反転した。だからこの候補は少数の日に乗っている」と書くことがある。本稿が点検するのは、著者自身がこれまで使ってきたこの検定である。上位を除けば、どんな列でも平均は必ず下がる。最大の観測を取り除くのだから、これは算術であって候補の脆さではない。したがって比べる相手が要る。自然な相...

Yuuki Yamagishi · 0 citations
#small language model Open access Sep 2026

The Maximum Drawdown Belongs to the Order of the Trades — Shuffling Alone Moves It by up to 4.23 Times Its Smallest Value, While the Total and the Win Rate Do Not Move at All [F031]

資金曲線を見せて「最大ドローダウンは N pips でした」と書くことがある。だが合計は足す順番に依らない。どう並べ替えても同じであり、それは足し算がそういうものだからで、測定の結果ではない。一方で最大ドローダウンは途中経過を見る量であり、負けが固まって来れば深くなり、散らばって来れば浅くなる。同じ取引の集合でも並べ方で変わる...

Yuuki Yamagishi · 0 citations

From tech blogs

See all →

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.