paper_id string | title string | published_date string | arxiv_categories list | legal_license string | license_url string | commercial_use_allowed bool | primary_author string | total_authors_count int64 | authors list | affiliations list | arxiv_url string | code_audit_status string | repo_urls list | repo_status_codes list | repo_verified_live list | abstract string | title_vector_384d list | abstract_vector_384d list | target_industry string | core_problem_addressed string | key_technical_innovation string | key_quantitative_result string | detected_technical_methods list | extracted_metrics list | trend_velocity_tier string | papers_last_30_days int64 | trl_score int64 | readiness_tier string | trl_justification string | audited_at string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2608.07463v1 | MirrorWorld: Taming Video Diffusion Models for Mirror Reflection Generation | 2026-08-07T17:58:10Z | [
"cs.CV",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Youjun Zhao | 4 | [
"Youjun Zhao",
"Alex Warren",
"Gary K. L. Tam",
"Rynson W. H. Lau"
] | [] | http://arxiv.org/abs/2608.07463v1 | NOT_DETECTED | [] | [] | [] | Recent advances in video diffusion models (VDMs) have enabled high-fidelity video synthesis. However, generating mirror reflections remains challenging because the content within a mirror must remain consistent with the surrounding scene. Existing VDMs are not specifically designed to model scene-to-mirror relationship... | [
0.025120999664068222,
-0.14161400496959686,
0.08216200023889542,
-0.060628000646829605,
0.02967200055718422,
-0.06446699798107147,
-0.05053500086069107,
-0.11977899819612503,
0.021846000105142593,
-0.043088000267744064,
-0.032641999423503876,
-0.009211000055074692,
-0.009882000274956226,
-... | [
0.04027700051665306,
-0.15169599652290344,
0.07557500153779984,
-0.014440000057220459,
0.04899799823760986,
-0.04700300097465515,
-0.09846299886703491,
-0.11840499937534332,
0.012507000006735325,
-0.05046800151467323,
-0.04699400067329407,
-0.008019999600946903,
-0.022923000156879425,
0.00... | LLM Fine-Tuning, Quantization & Model Optimization | Recent advances in video diffusion models (VDMs) have enabled high-fidelity video synthesis. | Existing VDMs are not specifically designed to model scene-to-mirror relationships, which can lead to reflections with incorrect content or inconsistent spatial arrangements. | Experimental results show that MirrorWorld achieves improved reflection reconstruction quality over representative image-based reflection generation methods and strong video inpainting baselines. | [
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:45.089318 |
2608.07430v1 | Diffusion LLMs as Targets and Adversaries: Mechanistic Safety Exploits | 2026-08-07T17:17:18Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Elena Dumitrescu | 4 | [
"Elena Dumitrescu",
"Gert Lek",
"Lydia Y. Chen",
"Jérémie Decouchant"
] | [] | http://arxiv.org/abs/2608.07430v1 | VERIFIED_LIVE | [
"https://github.com/ellyoana/sn-guided-diffusion"
] | [
200
] | [
true
] | Diffusion Large Language Models (DLLMs) replace autoregressive next-token prediction with iterative parallel denoising, yet their internal safety mechanisms remain poorly understood. In this work, we investigate DLLMs both as targets and as adversaries, exposing mechanistic vulnerabilities in diffusion-based alignment.... | [
0.02486399933695793,
-0.03345400094985962,
0.01714799925684929,
-0.030768999829888344,
0.07829199731349945,
-0.05929199978709221,
0.01344900019466877,
0.041019000113010406,
-0.014694999903440475,
0.07643599808216095,
0.02216400019824505,
0.04352499917149544,
0.07127899676561356,
-0.0126289... | [
-0.06247299909591675,
-0.11212000250816345,
-0.014240999706089497,
-0.0484049990773201,
0.036862000823020935,
0.017495999112725258,
-0.011742999777197838,
-0.009898999705910683,
0.0669659972190857,
0.01573999971151352,
0.029174000024795532,
-0.041832998394966125,
-0.022192999720573425,
-0.... | LLM Fine-Tuning, Quantization & Model Optimization | Diffusion Large Language Models (DLLMs) replace autoregressive next-token prediction with iterative parallel denoising, yet their internal safety mechanisms remain poorly understood. | We first show that safety alignment in DLLMs remains sparse and transferable across architectures. | Compared to prior jailbreaking frameworks, our method achieves competitive transferability with orders-of-magnitude lower generation cost. | [
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Benchmarks (+1) | 2026-08-10T16:37:46.700480 |
2608.07427v1 | A Picture is Worth a Thousand Tokens: How Vision Language Models Cut AI Energy Costs While Improving Accuracy | 2026-08-07T17:14:45Z | [
"cs.AI",
"cs.PF"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Bhavika Jalli | 3 | [
"Bhavika Jalli",
"Nikhil Korati Prasanna",
"Jayanta Choudhury"
] | [] | http://arxiv.org/abs/2608.07427v1 | NOT_DETECTED | [] | [] | [] | LLM inference accounts for over 90% of AI operational energy, scaling directly with input token count---a critical inefficiency for telecom network analytics and numerical time-series data analysis (NTSDA), where raw multivariate KPI windows from 4G/5G cell sites expand into thousands of floating-point tokens. Vision-L... | [
0.021236000582575798,
-0.04329700022935867,
0.028728000819683075,
0.0213019996881485,
0.08423499763011932,
0.0007229999755509198,
0.00546099990606308,
-0.014732999727129936,
0.09210299700498581,
-0.044218000024557114,
-0.061567001044750214,
-0.08103299885988235,
0.04002799838781357,
0.0052... | [
-0.01565299928188324,
-0.009453999809920788,
0.024470999836921692,
0.0191430002450943,
0.04635699838399887,
-0.02990500070154667,
-0.0404059998691082,
0.012841000221669674,
0.0862250030040741,
-0.005584999918937683,
-0.10395699739456177,
-0.020431000739336014,
0.025929000228643417,
-0.0260... | LLM Fine-Tuning, Quantization & Model Optimization | LLM inference accounts for over 90% of AI operational energy, scaling directly with input token count---a critical inefficiency for telecom network analytics and numerical time-series data analysis (NTSDA), where raw multivariate KPI windows from 4G/5G cell sites expand into thousands of floating-point tokens. | Vision-Language Models (VLMs) eliminate this mismatch by encoding time-series as 2D plots, achieving 3.6-10.4x input token reduction across Llama-3.2-90B, Qwen2.5-VL-72B, and Pixtral-12B architectures. | These results establish VLMs as an energy-efficient and accuracy-superior modality for numerical time-series workloads, providing empirical grounding for AI inference systems that treat energy consumption as a first-class engineering constraint. | [
"Large Language Model (LLM)",
"Anomaly Detection"
] | [
"20.6x improvement",
"F1 = 0.82"
] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:46.714727 |
2608.07287v1 | A foundation-model approach to pediatric headache classification from rs-fMRI | 2026-08-07T14:46:57Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Guilherme S. Imai Aldeia | 8 | [
"Guilherme S. Imai Aldeia",
"Clara Moon",
"Julie Shulman",
"Navil Sethna",
"Allison Smith",
"Alyssa Lebel",
"William G. La Cava",
"Scott Holmes"
] | [] | http://arxiv.org/abs/2608.07287v1 | NOT_DETECTED | [] | [] | [] | Headache is the most common neurological disorder in children and substantially affects quality of life. We investigated whether resting-state functional MRI (rs-fMRI) can support pediatric headache classification using machine learning. We encoded rs-fMRI data using NeuroSTORM, a recent foundation model, and fine-tune... | [
0.03170600160956383,
-0.07965599745512009,
0.008023999631404877,
0.03395399823784828,
0.014627000316977501,
0.062421999871730804,
-0.0004199999966658652,
-0.015786999836564064,
0.0164170004427433,
0.01869400031864643,
-0.023346999660134315,
-0.07015000283718109,
-0.0020749999675899744,
0.1... | [
-0.030774999409914017,
-0.09055200219154358,
-0.0049390001222491264,
0.08445700258016586,
0.022526999935507774,
0.072502002120018,
0.007662999909371138,
-0.0071749999187886715,
-0.0028409999795258045,
-0.024226000532507896,
-0.03776700049638748,
-0.05946800112724304,
-0.03779799863696098,
... | LLM Fine-Tuning, Quantization & Model Optimization | Headache is the most common neurological disorder in children and substantially affects quality of life. | We investigated whether resting-state functional MRI (rs-fMRI) can support pediatric headache classification using machine learning. | Results suggest that the approach can distinguish chronic migraine but has difficulty differentiating other headache subtypes from chronic migraine. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:46.721129 |
2608.07267v1 | WNM-3D: A World Navigation Model with 3D Scene Conditioning for Closed-Loop VLN | 2026-08-07T14:29:00Z | [
"cs.AI",
"cs.CV",
"cs.RO"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yuehao Huang | 10 | [
"Yuehao Huang",
"Yunzi Wu",
"Xiaotao Zhang",
"Xinhai Li",
"Jiankun Dong",
"Jiajun Lv",
"Chi Zhang",
"Chenjia Bai",
"Yong Liu",
"Xuelong Li"
] | [] | http://arxiv.org/abs/2608.07267v1 | NOT_DETECTED | [] | [] | [] | Recent vision-language navigation (VLN) systems increasingly adapt pretrained vision-language models (VLMs) into vision-language-action (VLA) policies that map egocentric observations and language instructions directly to navigation actions. Although semantically capable, such action-centric training does not explicitl... | [
0.025042999535799026,
-0.08729899674654007,
-0.027343999594449997,
-0.03255600109696388,
0.052319999784231186,
-0.017482999712228775,
-0.012690999545156956,
-0.04417699947953224,
-0.05212799832224846,
0.024535000324249268,
0.05502700060606003,
-0.014192000031471252,
0.048121001571416855,
-... | [
0.00698699988424778,
-0.12224400043487549,
0.018724000081419945,
-0.026940999552607536,
0.07225999981164932,
0.05111699923872948,
-0.006649000104516745,
-0.04810899868607521,
0.0037609999999403954,
-0.009424000047147274,
0.0005239999736659229,
-0.029217999428510666,
0.0034340000711381435,
... | LLM Fine-Tuning, Quantization & Model Optimization | Recent vision-language navigation (VLN) systems increasingly adapt pretrained vision-language models (VLMs) into vision-language-action (VLA) policies that map egocentric observations and language instructions directly to navigation actions. | Generative world-action models (WAMs) jointly predict future observations and actions, yet existing WAMs for continuous VLN do not condition joint future-view and action generation on geometry-aware representations inferred from the observed history. | On a fixed near-goal evaluation set, WNM-3D also achieves higher flow-action consistency and lower visual-motion error. | [
"Large Language Model (LLM)",
"Transformer Architecture",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:46.727549 |
2608.07226v1 | Dual-Node NVIDIA DGX Spark over Tailscale: A Remote-Access Testbed for Distributed LLM Training and Cyber-Threat-Intelligence Fine-Tuning | 2026-08-07T13:41:57Z | [
"cs.AR",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Vasanth Iyer | 1 | [
"Vasanth Iyer"
] | [
"NVIDIA Research"
] | http://arxiv.org/abs/2608.07226v1 | NOT_DETECTED | [] | [] | [] | Compact AI systems make local language-model experimentation increasingly accessible, yet practical evidence for multi-node training on desktop-class accelerators remains limited. This report presents a proof-of-concept deployment of distributed NanoChat pretraining across two NVIDIA DGX Spark systems, each with a GB10... | [
-0.013106999918818474,
-0.032565999776124954,
0.05651799961924553,
0.015162999741733074,
0.0745450034737587,
0.010835999622941017,
-0.057693000882864,
0.011896000243723392,
-0.036215998232364655,
0.04302600026130676,
-0.048344001173973083,
0.011320999823510647,
0.004563000053167343,
-0.033... | [
-0.014651999808847904,
-0.06687700003385544,
0.0034280000254511833,
0.04430599883198738,
0.015204999595880508,
-0.04475200176239014,
-0.05733000114560127,
0.036667000502347946,
-0.02711700089275837,
0.010706000030040741,
-0.04487999901175499,
-0.008820000104606152,
0.0035810000263154507,
-... | LLM Fine-Tuning, Quantization & Model Optimization | Compact AI systems make local language-model experimentation increasingly accessible, yet practical evidence for multi-node training on desktop-class accelerators remains limited. | This report presents a proof-of-concept deployment of distributed NanoChat pretraining across two NVIDIA DGX Spark systems, each with a GB10 Grace Blackwell system-on-chip and 128 GB of unified memory, administered remotely over a Tailscale mesh VPN and connected for training by a dedicated 200 Gb/s QSFP56 direct fiber... | CTI-specific categories improved while general-knowledge categories regressed, for a small overall change from 2.06 to 2.29 on a 0-10 scale. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:46.734480 |
2608.07208v1 | Measuring Concept Content in Text from LLM Activations: ESG Evidence from Concept Vectors and Linear Probes | 2026-08-07T13:21:50Z | [
"cs.CL",
"cs.AI",
"cs.LG",
"econ.GN"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Luc Hazenoot | 3 | [
"Luc Hazenoot",
"Zhaochun Ren",
"Amirhossein Zohrehvand"
] | [] | http://arxiv.org/abs/2608.07208v1 | NOT_DETECTED | [] | [] | [] | Existing measures of how much a text is about a concept read the surface of the text: dictionary word shares, topic proportions, embedding similarities. They score the words a text uses, not the judgment a reader forms about it. Recent work has shown that a gap exists in what Large Language Models (LLMs) know internall... | [
0.09413500130176544,
-0.09925299882888794,
0.035895999521017075,
0.025359999388456345,
0.09052299708127975,
-0.004296000115573406,
0.044684998691082,
-0.01283000037074089,
0.07958599925041199,
0.033684998750686646,
-0.05244600027799606,
0.02182300016283989,
0.009998999536037445,
-0.0107370... | [
0.07202000170946121,
-0.08388300240039825,
-0.07339499890804291,
0.053745001554489136,
0.09271100163459778,
-0.05422699823975563,
0.0702190026640892,
0.026170000433921814,
0.020445000380277634,
0.04641000181436539,
-0.06337500363588333,
0.029696999117732048,
0.023663999512791634,
0.0049140... | LLM Fine-Tuning, Quantization & Model Optimization | Existing measures of how much a text is about a concept read the surface of the text: dictionary word shares, topic proportions, embedding similarities. | This paper asks whether that internal knowledge, read by monitoring the activations of frozen, out-of-the-box LLMs, can stand in for task-specific fine-tuning when measuring concept content, and which extraction method reads it best. | We demonstrate the approach on financial text, a domain studied extensively and served by established annotated resources, using a human-annotated Environmental, Social and Governance (ESG) dataset. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:46.740621 |
2608.07193v1 | An AI4AI Framework for Visual Token Pruning | 2026-08-07T13:07:40Z | [
"cs.LG",
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Zhen Liu | 6 | [
"Zhen Liu",
"Wenli Huang",
"Wei Song",
"Yuhan Liu",
"Zhiqin Yang",
"Jingwen Fu"
] | [] | http://arxiv.org/abs/2608.07193v1 | NOT_DETECTED | [] | [] | [] | Visual-token pruning can substantially reduce the inference cost of multimodal large language models (MLLMs), yet existing methods largely rely on fixed, handcrafted heuristics and costly expert trial and error. As pruning objectives, budgets, and model architectures diversify, manually navigating the expanding design ... | [
0.009920000098645687,
0.021415000781416893,
0.0031059999018907547,
-0.027490999549627304,
0.0242490004748106,
-0.012888999655842781,
0.019923999905586243,
0.0027119999285787344,
0.009437999688088894,
-0.023081999272108078,
0.03430600091814995,
-0.04712200164794922,
0.049244001507759094,
0.... | [
0.04722699895501137,
-0.029326999559998512,
0.02536099962890148,
0.00955399964004755,
0.09941200166940689,
-0.005886000115424395,
-0.05110999941825867,
0.033709000796079636,
0.03543800115585327,
-0.019007999449968338,
-0.041586000472307205,
-0.0560929998755455,
0.057075001299381256,
0.0336... | LLM Fine-Tuning, Quantization & Model Optimization | Visual-token pruning can substantially reduce the inference cost of multimodal large language models (MLLMs), yet existing methods largely rely on fixed, handcrafted heuristics and costly expert trial and error. | As pruning objectives, budgets, and model architectures diversify, manually navigating the expanding design space becomes increasingly difficult. | Experiments on 14 multimodal benchmarks and three MLLM backbones demonstrate the effectiveness, efficiency, and transferability of AutoPrune. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:46.747550 |
2608.07169v1 | Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory | 2026-08-07T12:43:00Z | [
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Taeil Kim | 3 | [
"Taeil Kim",
"Kangsan Kim",
"Sung Ju Hwang"
] | [] | http://arxiv.org/abs/2608.07169v1 | NOT_DETECTED | [] | [] | [] | Memory systems have shown promise for improving agent performance, but their potential remains largely unexplored for small language models, which struggle to generate sufficient successful trajectories on their own. We propose Agent Memory Distillation (AMD), a training-free framework that transfers structured knowled... | [
0.08422800153493881,
-0.054965000599622726,
0.02548000030219555,
-0.009012999944388866,
-0.004894999787211418,
0.0018090000376105309,
-0.03133599832653999,
-0.009843000210821629,
-0.0031950001139193773,
-0.0015780000248923898,
0.008138000033795834,
0.04728199914097786,
0.055424999445676804,
... | [
0.04518600180745125,
-0.03855299949645996,
0.018884999677538872,
0.015641000121831894,
-0.0427360013127327,
-0.04969500005245209,
-0.025956999510526657,
0.025665000081062317,
0.03041600063443184,
-0.030510999262332916,
-0.04433000087738037,
0.02823599986732006,
0.03930100053548813,
0.03579... | LLM Fine-Tuning, Quantization & Model Optimization | Memory systems have shown promise for improving agent performance, but their potential remains largely unexplored for small language models, which struggle to generate sufficient successful trajectories on their own. | We propose Agent Memory Distillation (AMD), a training-free framework that transfers structured knowledge from a large teacher agent to a small student agent through hierarchical memory. | We evaluate AMD on three tool-use benchmarks using four student models (4B-8B parameters) with GPT-5-mini as the teacher, achieving average accuracy gains of 27.2%p, 11.2%p, and 3.4%p on AppWorld, BFCL V3, and ToolSandbox, while consistently outperforming existing memory-based baselines. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:46.751236 |
2608.07126v1 | PHOENIX: Fine-Tuned SLM-Powered Autonomous Satellite Lifetime Extension via Predictive Self-Healing and Multi-Agent AI Recovery | 2026-08-07T11:38:24Z | [
"cs.HC",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Sumaiya Islam | 2 | [
"Sumaiya Islam",
"Harsha Kumara Moraliyage"
] | [] | http://arxiv.org/abs/2608.07126v1 | NOT_DETECTED | [] | [] | [] | Most CubeSats, small and low-cost satellites roughly the size of a shoebox, do not survive as long as they were designed to: a study of 178 missions found that only 48-65% remain operational after two years, against a designed lifetime of 2-5 years. The deeper issue is that a CubeSat in low Earth orbit (LEO) is physica... | [
-0.059401001781225204,
-0.04707799851894379,
0.04245999827980995,
0.047269001603126526,
0.024268999695777893,
-0.056095998734235764,
0.009322999976575375,
-0.09926799684762955,
-0.051274001598358154,
0.04534599930047989,
-0.01759600080549717,
0.042086001485586166,
-0.001373999984934926,
-0... | [
0.003091000020503998,
-0.08037299662828445,
0.045740000903606415,
0.06769999861717224,
0.04055099934339523,
-0.0568699985742569,
0.022231999784708023,
0.028178999200463295,
-0.03239399939775467,
0.03901999816298485,
-0.07839000225067139,
0.04926599934697151,
0.009662000462412834,
-0.006370... | LLM Fine-Tuning, Quantization & Model Optimization | Most CubeSats, small and low-cost satellites roughly the size of a shoebox, do not survive as long as they were designed to: a study of 178 missions found that only 48-65% remain operational after two years, against a designed lifetime of 2-5 years. | We propose PHOENIX (Predictive Health On-orbit Edge Neural Intelligence eXtension) to give the satellite its own fault reasoning capability. | We report preliminary results on the ESA Anomaly Detection Benchmark (14 years, 76 channels, 118 labeled faults). | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Multi-Agent System",
"Anomaly Detection"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:46.755586 |
2608.07088v1 | RoRA: Role-Oriented Regional Allocation for Visual Token Pruning in MLLMs | 2026-08-07T10:39:47Z | [
"cs.CV",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Qiyanhui Lu | 9 | [
"Qiyanhui Lu",
"Han Wu",
"Rongjian Xu",
"Tingzhang Luo",
"Cheng Fan",
"Xinghao Chen",
"Minjing Dong",
"Jufeng Yang",
"Jianyuan Guo"
] | [
"NVIDIA Research"
] | http://arxiv.org/abs/2608.07088v1 | NOT_DETECTED | [] | [] | [] | Multimodal large language models (MLLMs) encode images as long visual token sequences, making prefilling and KV-cache storage expensive. Existing training-free pruning methods select tokens by importance, diversity, or spatial coverage, but treat retained tokens as interchangeable and do not explicitly track which obje... | [
0.10138999670743942,
-0.06266400218009949,
0.006864000111818314,
-0.0025919999461621046,
0.05445599928498268,
0.010699000209569931,
-0.04909300059080124,
-0.05122800171375275,
-0.000999000039882958,
-0.021463999524712563,
0.005985000170767307,
0.013969999738037586,
0.06426899880170822,
-0.... | [
0.028627999126911163,
-0.031766001135110855,
-0.008081999607384205,
0.035043999552726746,
0.1667719930410385,
0.023979000747203827,
-0.080765001475811,
-0.032513998448848724,
0.045278001576662064,
-0.04867099970579147,
-0.005212000105530024,
-0.03166000172495842,
0.022061999887228012,
0.06... | LLM Fine-Tuning, Quantization & Model Optimization | Multimodal large language models (MLLMs) encode images as long visual token sequences, making prefilling and KV-cache storage expensive. | Existing training-free pruning methods select tokens by importance, diversity, or spatial coverage, but treat retained tokens as interchangeable and do not explicitly track which object-related regions are already covered. | At a 66.7% pruning ratio, RoRA requires only 0.7 ms for token selection and reduces end-to-end inference time by 24.6%, corresponding to a 1.33x speedup over unpruned inference on an NVIDIA H800. | [
"Small Language Model (SLM)"
] | [
"1.33x speedup"
] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:46.762044 |
2608.07053v1 | Unsupervised Adaptation of PDE Foundation Models | 2026-08-07T10:01:50Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Ziye Song | 5 | [
"Ziye Song",
"Zhao Wei",
"Xin Yu",
"Ivor Tsang",
"Yueming Lyu"
] | [] | http://arxiv.org/abs/2608.07053v1 | NOT_DETECTED | [] | [] | [] | Pretrained partial differential equation (PDE) foundation models can generalize across different equations, but adapting them to unseen PDE systems typically requires dense solution data, which is often expensive or unavailable. To address this limitation, we propose an unsupervised PDE-based finetuning framework that ... | [
-0.04377000033855438,
-0.08916900306940079,
0.0851529985666275,
0.03009199909865856,
0.13663500547409058,
0.02559100091457367,
-0.07964000105857849,
-0.039698999375104904,
-0.00572500005364418,
-0.03394800052046776,
0.007375000044703484,
-0.06487900018692017,
0.004168999847024679,
-0.01492... | [
-0.02164600044488907,
-0.1133359968662262,
0.12225300073623657,
0.05118799954652786,
0.11404299736022949,
0.011447999626398087,
-0.12053100019693375,
-0.043807998299598694,
-0.01658499985933304,
-0.05925200134515762,
-0.022864000871777534,
-0.061994001269340515,
-0.06562700122594833,
0.063... | LLM Fine-Tuning, Quantization & Model Optimization | Pretrained partial differential equation (PDE) foundation models can generalize across different equations, but adapting them to unseen PDE systems typically requires dense solution data, which is often expensive or unavailable. | To address this limitation, we propose an unsupervised PDE-based finetuning framework that eliminates the need for ground-truth solutions. | Our method achieves performance comparable to supervised LoRA finetuning without requiring any ground-truth solutions, while consistently outperforming competitive neural operator baselines and recent PDE foundation models across heterogeneous PDE benchmarks spanning multiple spatial dimensions. | [
"Transformer Architecture",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:46.767710 |
2608.06901v1 | Prune Once: Retraining-Free Task-Agnostic Pruning for Vision-Language Models | 2026-08-07T07:36:19Z | [
"cs.CV",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Minseok Kang | 6 | [
"Minseok Kang",
"Hyunwoo Kim",
"Chanyoung Kim",
"Minwoo Kim",
"Jaekoo Lee",
"Dahuin Jung"
] | [] | http://arxiv.org/abs/2608.06901v1 | VERIFIED_LIVE | [
"https://github.com/cau-hai-lab/PORTA.git"
] | [
200
] | [
true
] | Vision-language models (VLMs) have achieved remarkable generalization across diverse multimodal tasks through large-scale pre-training, yet their rapidly increasing computational and memory requirements pose significant challenges for deployment in constrained environments. Existing pruning strategies often depend on t... | [
0.008306000381708145,
-0.07049000263214111,
0.02281999960541725,
0.0033919999841600657,
0.08007100224494934,
0.07445400208234787,
0.017544999718666077,
-0.041186001151800156,
0.02942799963057041,
-0.06442700326442719,
0.05163700133562088,
-0.048815999180078506,
0.058552999049425125,
0.0433... | [
0.04311300069093704,
-0.05473100021481514,
-0.012676999904215336,
-0.015765000134706497,
0.12014999985694885,
0.005934999790042639,
-0.03881999850273132,
-0.05299000069499016,
-0.017495999112725258,
-0.08711300045251846,
-0.004007999785244465,
-0.06306400150060654,
-0.012338999658823013,
0... | LLM Fine-Tuning, Quantization & Model Optimization | Vision-language models (VLMs) have achieved remarkable generalization across diverse multimodal tasks through large-scale pre-training, yet their rapidly increasing computational and memory requirements pose significant challenges for deployment in constrained environments. | We introduce a retraining-free VLM pruning framework called PORTA that derives a task- and modality-agnostic importance formulation based on activation variation, estimated from generic calibration data, which reliably captures feature-level representation utility across modalities. | Extensive experiments across VLM architectures, such as CLIP, BLIP, and Qwen2-VL, demonstrate that PORTA achieves competitive downstream performance under high sparsity without requiring any retraining, supporting efficient VLM compression. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Benchmarks (+1) | 2026-08-10T16:37:48.125408 |
2608.06795v1 | LoRAScan: Detecting Backdoor Prompts in Low-Rank Adapters for Large Language Models via Down-Projection Activation Spikes | 2026-08-07T04:36:24Z | [
"cs.CR",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Doniyorkhon Obidov | 4 | [
"Doniyorkhon Obidov",
"Honggang Yu",
"Xiaolong Guo",
"Kaichen Yang"
] | [] | http://arxiv.org/abs/2608.06795v1 | NOT_DETECTED | [] | [] | [] | Low-rank adaptation (LoRA) enables efficient specialization and distribution of large language models through compact adapters. However, untrusted adapters introduce a supply-chain threat: a backdoored adapter can cause a model to generate harmful content, malicious code, political propaganda, or covert advertisements ... | [
-0.04849199950695038,
-0.08944299817085266,
0.044509999454021454,
-0.010850000195205212,
0.09873399883508682,
0.04670200124382973,
-0.023990999907255173,
-0.037539999932050705,
0.057130999863147736,
-0.04795200005173683,
-0.028686000034213066,
-0.054503001272678375,
0.05381700024008751,
-0... | [
-0.06823600083589554,
-0.0732249990105629,
-0.008821000345051289,
-0.008856999687850475,
0.12759500741958618,
0.030841000378131866,
-0.01839599944651127,
-0.019682999700307846,
0.03022499941289425,
-0.059627000242471695,
0.021971000358462334,
-0.00802099984139204,
0.11867699772119522,
-0.0... | LLM Fine-Tuning, Quantization & Model Optimization | Low-rank adaptation (LoRA) enables efficient specialization and distribution of large language models through compact adapters. | However, untrusted adapters introduce a supply-chain threat: a backdoored adapter can cause a model to generate harmful content, malicious code, political propaganda, or covert advertisements when an input contains a hidden trigger. | Across standard LLM backdoor benchmarks, LoRAScan rejects approximately 98.49 of malicious inputs with a small error rate on clean inputs, outperforming existing defenses across diverse evaluation settings. | [
"Large Language Model (LLM)",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:48.153037 |
2608.06792v1 | Progressive Alignment of Recommender Foundation Model through Multi-Phase Post-Training | 2026-08-07T04:30:34Z | [
"cs.IR",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Oseong Choi | 5 | [
"Oseong Choi",
"Hoeinn Kim",
"Jihoon Lee",
"Byungsoo Kang",
"Taeyeong Jang"
] | [] | http://arxiv.org/abs/2608.06792v1 | VERIFIED_LIVE | [
"https://github.com/webtoon/rec-fm-progressive-alignment"
] | [
200
] | [
true
] | Foundation model(FM) for recommendation has shown strong ability to model long-horizon sequential user behavior. In practice, a single pretrained foundation model is often adapted to diverse downstream serving surfaces through Supervised Fine-Tuning(SFT). However, optimizing task-specific objectives such as clicks or l... | [
-0.009550999850034714,
-0.10531000047922134,
-0.05548999831080437,
0.0015330000314861536,
0.08184699714183807,
0.05174500122666359,
0.011083000339567661,
-0.0007270000060088933,
-0.019739000126719475,
-0.06561700254678726,
0.0006040000007487833,
0.05234900116920471,
0.07460500299930573,
0.... | [
-0.028807999566197395,
-0.16224700212478638,
-0.008747999556362629,
0.021260999143123627,
0.0030950000509619713,
0.05830200016498566,
0.03252600133419037,
-0.013794000260531902,
-0.0015589999966323376,
-0.05591399967670441,
-0.0895249992609024,
0.05916899815201759,
-0.025326000526547432,
0... | LLM Fine-Tuning, Quantization & Model Optimization | Foundation model(FM) for recommendation has shown strong ability to model long-horizon sequential user behavior. | We propose a three-phase progressive post-training framework that explicitly separates downstream adaptation from business-metric alignment. | Large-scale online A/B tests further show that the proposed framework improves production recommendation quality over a conventional non-foundation baseline. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Benchmarks (+1) | 2026-08-10T16:37:48.997462 |
2608.06778v1 | Retrieval-Constrained Policy Optimization for Attack Technique Extraction from Cyber Threat Intelligence | 2026-08-07T03:52:33Z | [
"cs.CR",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Jiayun Zhang | 4 | [
"Jiayun Zhang",
"Junshen Xu",
"Zejun Xie",
"Yi Fan"
] | [] | http://arxiv.org/abs/2608.06778v1 | NOT_DETECTED | [] | [] | [] | Mapping cyber threat intelligence (CTI) text to MITRE ATT&CK techniques is essential for structured threat analysis, yet manual annotation is costly and does not scale. The ATT&CK taxonomy comprises several hundred attack techniques, and a single CTI passage may describe multiple techniques, making accurate and complet... | [
-0.05258699879050255,
0.004790999926626682,
-0.03579799830913544,
0.021000999957323074,
0.046303000301122665,
0.1003979966044426,
0.06747200340032578,
-0.054416999220848083,
0.013485999777913094,
-0.015393000096082687,
-0.011377999559044838,
-0.009734000079333782,
0.10033699870109558,
0.00... | [
-0.07800199836492538,
0.0007399999885819852,
-0.04401199892163277,
-0.024318000301718712,
0.05128899961709976,
0.09882599860429764,
0.028373999521136284,
0.009548000060021877,
-0.044234998524188995,
-0.037264999002218246,
-0.04822099953889847,
-0.08426599949598312,
0.0787070021033287,
0.02... | LLM Fine-Tuning, Quantization & Model Optimization | Mapping cyber threat intelligence (CTI) text to MITRE ATT&CK techniques is essential for structured threat analysis, yet manual annotation is costly and does not scale. | The ATT&CK taxonomy comprises several hundred attack techniques, and a single CTI passage may describe multiple techniques, making accurate and complete extraction challenging. | Across four CTI benchmarks, TTP-R1 achieves the best average F1, improving sub-technique-level F1 by 7.4 percentage points over Claude Sonnet 4.5 with retrieval augmentation, while running 28x faster when served as an 8B-parameter model on a single GPU. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)"
] | [
"28x faster"
] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:49.021825 |
2608.06763v1 | CubicQuant: Parametric Non-Uniform Codebooks for High-Throughput LLM Inference with 1-8-Bit Weights | 2026-08-07T03:36:07Z | [
"cs.LG",
"cs.DC"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Xuetian Gao | 1 | [
"Xuetian Gao"
] | [] | http://arxiv.org/abs/2608.06763v1 | NOT_DETECTED | [] | [] | [] | Weight quantization for large-language-model inference must balance adaptive reconstruction levels with representations regular enough for efficient GPU execution. Uniform integers constrain each group to a linear grid. Low-bit floating-point formats use a fixed exponent-mantissa structure, while learned codebooks gain... | [
-0.03234200179576874,
-0.13743999600410461,
-0.06078900024294853,
0.005890999920666218,
-0.004631000105291605,
-0.05677900090813637,
-0.08492500334978104,
0.0003699999942909926,
-0.025953000411391258,
-0.030987000092864037,
-0.010529999621212482,
-0.0524739995598793,
-0.000539999979082495,
... | [
-0.07186300307512283,
-0.024636000394821167,
-0.07381699979305267,
0.03005100041627884,
-0.014666000381112099,
-0.04428999871015549,
-0.1045060008764267,
-0.02576499991118908,
0.0024610001128166914,
-0.06676600128412247,
-0.07176800072193146,
-0.0203079991042614,
-0.042628999799489975,
0.0... | LLM Fine-Tuning, Quantization & Model Optimization | Weight quantization for large-language-model inference must balance adaptive reconstruction levels with representations regular enough for efficient GPU execution. | We introduce CubicQuant, a parametric non-uniform scalar format that preserves a dense integer code stream while adapting reconstruction levels within each weight group. | The results establish the format's representational promise and direct executability; downstream model quality and cross-device end-to-end performance remain open evaluation questions. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Quantization"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:37:49.048390 |
2608.06756v1 | Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence | 2026-08-07T03:24:25Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Ying Chen | 10 | [
"Ying Chen",
"Weizhen Li",
"Zhe Hu",
"Zhenjiang Li",
"Rui Jiang",
"Zhifeng Gu",
"Lihuang Fang",
"Jiangping Liu",
"Lei Yi",
"Jie Chen"
] | [] | http://arxiv.org/abs/2608.06756v1 | NOT_DETECTED | [] | [] | [] | Vision-language models are increasingly serving as the reasoning core of embodied agents. Robot execution is inherently iterative: each action reshapes the scene and physical state, continually renewing what must be perceived, reasoned about, and verified. Meeting these demands requires complementary capabilities that ... | [
0.044266000390052795,
-0.005303000099956989,
0.038155000656843185,
-0.04826800152659416,
0.03507800027728081,
0.007786999922245741,
0.07158199697732925,
0.004557999782264233,
0.01027899980545044,
0.032464999705553055,
-0.034963998943567276,
-0.0785600021481514,
0.01898300088942051,
0.02728... | [
0.07364799827337265,
-0.07967499643564224,
0.033323001116514206,
-0.018147999420762062,
0.060795001685619354,
0.011390999890863895,
0.026247000321745872,
-0.03983699902892113,
-0.007596000097692013,
0.014924000017344952,
-0.0177839994430542,
-0.07936400175094604,
0.02474600076675415,
0.123... | LLM Fine-Tuning, Quantization & Model Optimization | Vision-language models are increasingly serving as the reasoning core of embodied agents. | Existing approaches typically develop these capabilities against isolated, task-specific objectives, leaving open how they should be organized and integrated around execution as a whole. | Capek 0.5 improves the large majority of matched benchmark rows over its initialization, retains all four specialized capabilities in one checkpoint with quantified losses, and transfers to closed-loop embodied task execution. | [
"Reinforcement Learning (RL)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:49.060349 |
2608.06690v1 | Policy-Masked Private Experts: Auditable and Reversible Capability Access Control in Sparse MoE Models | 2026-08-07T01:39:36Z | [
"cs.CR",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Zhuoheng Huang | 2 | [
"Zhuoheng Huang",
"Mukesh Singh"
] | [] | http://arxiv.org/abs/2608.06690v1 | NOT_DETECTED | [] | [] | [] | Most language-model access controls regulate behavior while leaving the same computation available to every request. We study a different systems question: can trusted authorization determine which newly trained parameters are reachable by the forward pass? Policy-Masked Private Experts freezes a pretrained sparse Mixt... | [
-0.02878599986433983,
0.009390000253915787,
-0.02333500050008297,
0.045729998499155045,
-0.00863800011575222,
0.021278999745845795,
0.029563000425696373,
-0.07179299741983414,
-0.03458800166845322,
0.029798999428749084,
-0.0076489998027682304,
0.038488999009132385,
0.020754000172019005,
-0... | [
-0.0926859974861145,
-0.008721999824047089,
0.0054259998723864555,
0.04984999820590019,
0.028655000030994415,
-0.02418299950659275,
-0.006105999927967787,
-0.04916299879550934,
0.017235999926924706,
0.02853499911725521,
0.010882000438869,
-0.015992000699043274,
0.05471799895167351,
-0.0295... | LLM Fine-Tuning, Quantization & Model Optimization | Most language-model access controls regulate behavior while leaving the same computation available to every request. | We study a different systems question: can trusted authorization determine which newly trained parameters are reachable by the forward pass? | These results support auditable, reversible control over a trained parameter path, while showing that useful transfer remains distribution dependent. | [
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:49.070257 |
2608.06632v1 | Shape Your Feed: An LLM-based Agentic System for Conversational Recommendation | 2026-08-06T22:54:16Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Ziyun Xu | 13 | [
"Ziyun Xu",
"Bosen Ding",
"Yue Zhang",
"Ji Qi",
"Qingyuan Song",
"Jizhou Huang",
"Liwei Wang",
"Jefferey Santelli",
"Yue Weng",
"Qichao Que",
"Zhenheng Yang",
"Junfeng Pan",
"Linhong Zhu"
] | [] | http://arxiv.org/abs/2608.06632v1 | NOT_DETECTED | [] | [] | [] | Industrial recommendation systems predominantly adopt a passive ranking paradigm that infers user preferences from implicit behavioral signals (e.g., clicks, dwell time) rather than explicit, natural language inputs. As a result, users experience a persistent discrepancy between their explicit interests and what passiv... | [
-0.04466500133275986,
-0.11957400292158127,
-0.0006019999855197966,
0.005636999849230051,
0.0411049984395504,
-0.01666799932718277,
0.048680998384952545,
0.029639000073075294,
-0.020979000255465508,
-0.08177299797534943,
-0.051669999957084656,
0.011443999595940113,
0.036757998168468475,
0.... | [
0.0007340000011026859,
-0.1392659991979599,
-0.04030900076031685,
-0.014467000029981136,
0.07888299971818924,
0.01396000012755394,
0.04338899999856949,
0.014770000241696835,
-0.05090700089931488,
-0.04080500081181526,
-0.055580999702215195,
0.006250999867916107,
0.020434999838471413,
0.019... | LLM Fine-Tuning, Quantization & Model Optimization | Industrial recommendation systems predominantly adopt a passive ranking paradigm that infers user preferences from implicit behavioral signals (e.g., clicks, dwell time) rather than explicit, natural language inputs. | As a result, users experience a persistent discrepancy between their explicit interests and what passive behavioral algorithms deliver, limiting their ability to express nuanced preferences or steer their feed in real time. | Large-scale online A/B experiments on production traffic further demonstrate that SYF improves feed relevance and user sentiment, indicating a practical and scalable path toward interactive, user-steerable recommendation in industrial settings. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Multi-Agent System"
] | [
"98.85% accuracy"
] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:49.079688 |
2608.06630v1 | The Sparsity Whisperer | 2026-08-06T22:37:26Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Linghao Kong | 6 | [
"Linghao Kong",
"Inimai Subramanian",
"Micah Adler",
"Dan Alistarh",
"Dan Gutfreund",
"Nir Shavit"
] | [] | http://arxiv.org/abs/2608.06630v1 | NOT_DETECTED | [] | [] | [] | Pruning reduces the inference cost of large language models, but existing criteria primarily preserve large activations or reconstruct layer outputs. We argue that this overlooks a key computation performed by particularly sparsity-sensitive neurons in the MLP up and gate projections: separating similar inputs into dis... | [
-0.07850799709558487,
-0.08971700072288513,
0.03888000175356865,
-0.030081000179052353,
-0.0046720001846551895,
-0.08960700035095215,
0.1031540036201477,
-0.05840799957513809,
0.024675000458955765,
-0.06132400035858154,
0.03200000151991844,
0.017938999459147453,
-0.01307700015604496,
-0.00... | [
-0.06246799975633621,
-0.11020699888467789,
0.026045000180602074,
0.024557000026106834,
0.0497249998152256,
0.0424249991774559,
-0.03342700004577637,
-0.0396450012922287,
0.0553009994328022,
-0.0511149987578392,
-0.004141999874264002,
0.03208800032734871,
-0.02696700021624565,
-0.024570999... | LLM Fine-Tuning, Quantization & Model Optimization | Pruning reduces the inference cost of large language models, but existing criteria primarily preserve large activations or reconstruct layer outputs. | We introduce a family of difference-informed pruning methods built upon this principle. | These results suggest that preserving output differences is a broadly useful and composable signal for post-training LLM sparsification. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:49.088465 |
2608.06609v1 | Automated item evaluation: Predicting item acceptance and rejection using LLM-generated critiques | 2026-08-06T21:52:18Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Hotaka Maeda | 2 | [
"Hotaka Maeda",
"Yikai Lu"
] | [] | http://arxiv.org/abs/2608.06609v1 | NOT_DETECTED | [] | [] | [] | Automated item evaluation (AIE) refers to the use of computational methods to assess item quality without requiring manual expert review or field testing of the items under evaluation. We aimed to build a near-comprehensive AIE model by predicting item acceptance and rejection from item text using historical rejection ... | [
-0.0320109985768795,
0.07097599655389786,
0.0007110000005923212,
0.03135399892926216,
0.11993999779224396,
-0.022898999974131584,
0.08394599705934525,
0.039326999336481094,
-0.006585000082850456,
0.012889999896287918,
-0.01774599961936474,
-0.04356599971652031,
0.08854799717664719,
-0.0317... | [
-0.06986600160598755,
0.017549000680446625,
-0.02432199940085411,
0.03897000104188919,
0.05923999845981598,
-0.030724000185728073,
0.04619399830698967,
0.05592700093984604,
0.021762000396847725,
0.06216000020503998,
-0.03623199835419655,
-0.02464599907398224,
0.09112299978733063,
-0.024553... | LLM Fine-Tuning, Quantization & Model Optimization | Automated item evaluation (AIE) refers to the use of computational methods to assess item quality without requiring manual expert review or field testing of the items under evaluation. | We fine-tuned a DeBERTaV3-large classifier on raw item text, a second DeBERTa classifier on Qwen3-generated item critiques, and a fusion model combining representations from both. | Incorporating item critiques alongside raw item text improved performance across most rejection reasons. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:49.099659 |
2608.06526v1 | GRASP: Reinforcing Language Model Anonymizers with Group Relative Policy Optimization | 2026-08-06T19:12:57Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Sajjad Ghiasvand | 2 | [
"Sajjad Ghiasvand",
"Nader Sehatbakhsh"
] | [] | http://arxiv.org/abs/2608.06526v1 | NOT_DETECTED | [] | [] | [] | Large language models can infer sensitive personal attributes, such as age, location, and occupation, from ordinary text, turning everyday writing into a privacy risk. Adversarial anonymization defends against this by rewriting a text with a capable language model that also plays the attacker, but it needs a powerful m... | [
-0.07607299834489822,
-0.08261100202798843,
-0.0042420001700520515,
-0.04703700169920921,
0.007627000100910664,
0.018350999802350998,
0.07822000235319138,
-0.08774100244045258,
0.07312899827957153,
-0.04541600123047829,
0.07474300265312195,
-0.06861600279808044,
0.09720800071954727,
0.0007... | [
-0.0909850001335144,
-0.016550999134778976,
0.04038500040769577,
0.02383499965071678,
0.04063500091433525,
0.04691600054502487,
0.03566399961709976,
-0.028031000867486,
0.03573499992489815,
0.011985000222921371,
0.04188999906182289,
0.03294200077652931,
0.039097998291254044,
-0.01882299967... | LLM Fine-Tuning, Quantization & Model Optimization | Large language models can infer sensitive personal attributes, such as age, location, and occupation, from ordinary text, turning everyday writing into a privacy risk. | We introduce \textbf{GRASP} (\textbf{G}roup-\textbf{R}elative \textbf{A}nonymization via \textbf{S}elf-refinement \textbf{P}olicy-optimization), which reinforces the local anonymizer online with Group Relative Policy Optimization. | Against adversarial anonymization driven by frontier models such as Gemini~2.5~Flash and Claude, it achieves a comparable or better overall trade-off while removing substantially more private information, and it runs entirely on-device at roughly $1\%$ of the GPT-4o teacher's cost. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:49.109693 |
2608.06471v1 | CyberForge: Verified Vulnerability Injection at Repository Level for Cybersecurity Agent Training | 2026-08-06T18:03:27Z | [
"cs.CR",
"cs.AI",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Amine Lbath | 7 | [
"Amine Lbath",
"Manan Suri",
"Aurelien Delaitre",
"Vadim Okun",
"Massih-Reza Amini",
"Ram D. Sriram",
"Dinesh Manocha"
] | [] | http://arxiv.org/abs/2608.06471v1 | NOT_DETECTED | [] | [] | [] | Despite recent advances, frontier large language model (LLM) agents remain limited in discovering and patching complex vulnerabilities in real-world software. Generally available agents can already aid attackers, who only need to find one exploitable weakness, while defenders must continuously identify and patch all vu... | [
-0.020880000665783882,
-0.014096000231802464,
-0.10444500297307968,
0.032280001789331436,
-0.0009730000165291131,
0.030275000259280205,
0.06763400137424469,
-0.048622000962495804,
-0.06641499698162079,
-0.05335799977183342,
0.02775599993765354,
-0.05903000012040138,
0.12092100083827972,
0.... | [
-0.060502998530864716,
-0.043758999556303024,
-0.06440900266170502,
-0.028857000172138214,
0.01014699973165989,
0.014333000406622887,
-0.01663699932396412,
0.004215999972075224,
-0.05748799815773964,
-0.015200000256299973,
-0.004064999986439943,
-0.07011199742555618,
0.0643170028924942,
0.... | LLM Fine-Tuning, Quantization & Model Optimization | Despite recent advances, frontier large language model (LLM) agents remain limited in discovering and patching complex vulnerabilities in real-world software. | We present CyberForge, a framework that synthesizes executable, repository-level security training data by injecting vulnerabilities into real C/C++ projects. | These gains generalize out of distribution to PatchEval, a corpus containing other programming languages, where every configuration also improves and the 31B student passes its teacher. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:49.119845 |
2608.06259v1 | RxnCLF: Contrastive Transformation-Aware Reaction Foundation Model for Improved Reactivity Prediction | 2026-08-06T16:51:23Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yiting Zheng | 4 | [
"Yiting Zheng",
"Cheng Fang",
"Anthony Donofrio",
"Haote Li"
] | [] | http://arxiv.org/abs/2608.06259v1 | NOT_DETECTED | [] | [] | [] | Reaction yield prediction remains challenging because labeled data are scarce and reaction space is both combinatorially large and sparsely populated, limiting the generalization of existing reaction representations. String-, fingerprint-, and graph-based reaction encodings only partially capture chemical transformatio... | [
0.0009730000165291131,
-0.1378909945487976,
-0.0001880000054370612,
0.05169599875807762,
0.1340479999780655,
0.04818100109696388,
-0.05707800015807152,
0.054343000054359436,
-0.01145000010728836,
0.03829900175333023,
-0.06605499982833862,
-0.07503700256347656,
-0.0018949999939650297,
0.066... | [
-0.05184299871325493,
-0.0821480005979538,
-0.025117000564932823,
0.06026700139045715,
0.08470699936151505,
0.06213900074362755,
-0.05712199956178665,
0.021764999255537987,
-0.012343999929726124,
-0.02473899908363819,
-0.04673200100660324,
-0.09902899712324142,
-0.003295999951660633,
0.055... | LLM Fine-Tuning, Quantization & Model Optimization | Reaction yield prediction remains challenging because labeled data are scarce and reaction space is both combinatorially large and sparsely populated, limiting the generalization of existing reaction representations. | We propose reaction contrastive learning foundation (RxnCLF), a self-supervised contrastive framework for reaction representation learning. | Our results highlight the promise of CRG-based RxnCLF as a scalable reaction foundation model, with the potential to generalize across broader reaction spaces and support diverse downstream reaction informatics tasks, including regioselectivity prediction, enantioselectivity prediction, and reaction condition optimizat... | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:49.130639 |
2608.06253v1 | MetaboLLM: a metabolomics-specialized large language model for biochemical knowledge integration and predictive metabolite graph construction | 2026-08-06T16:42:34Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Dohyun Ku | 4 | [
"Dohyun Ku",
"Min Gu Kwak",
"Francisco J. Pasquel",
"Jing Li"
] | [] | http://arxiv.org/abs/2608.06253v1 | NOT_DETECTED | [] | [] | [] | Metabolomics knowledge is distributed across heterogeneous resources and remains difficult to translate into predictive representations. We developed MetaboLLM, a metabolomics-specialized large language model adapted through continual pretraining, supervised fine-tuning, and structured retrieval, together with MetaboLL... | [
-0.0003549999964889139,
-0.04923500120639801,
-0.02971700020134449,
0.0008229999803006649,
-0.0032969999592751265,
0.01447600033134222,
-0.051931001245975494,
0.15061800181865692,
0.009538999758660793,
-0.010406999848783016,
0.019652999937534332,
-0.07198300212621689,
-0.0038479999639093876,... | [
-0.013558000326156616,
-0.053123001009225845,
-0.027156999334692955,
0.06024400144815445,
0.012225000187754631,
0.008441000245511532,
-0.07480499893426895,
0.1258459985256195,
-0.012494999915361404,
-0.05440700054168701,
-0.05739400163292885,
-0.03789500147104263,
-0.022577999159693718,
0.... | LLM Fine-Tuning, Quantization & Model Optimization | Metabolomics knowledge is distributed across heterogeneous resources and remains difficult to translate into predictive representations. | We developed MetaboLLM, a metabolomics-specialized large language model adapted through continual pretraining, supervised fine-tuning, and structured retrieval, together with MetaboLLM-GIN, which converts generated biochemical descriptions into metabolite graphs for patient-level prediction using a graph isomorphism ne... | These results show that domain-specialized language models can organize heterogeneous biochemical knowledge into predictive and interpretable metabolite graph representations. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:49.139564 |
2608.06243v2 | DASH: Divergence-Adaptive Supervision Horizons for On-Policy Self-Distillation of Reasoning Models | 2026-08-06T16:29:24Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | ZhiYan Hou | 12 | [
"ZhiYan Hou",
"Xinyu Tang",
"Hongyan An",
"Jianjin Zhang",
"Weizhen Wang",
"Yunyun Han",
"Gengsheng Li",
"Xiangzhao Hao",
"Haiyun Guo",
"Wenbin Hu",
"Jinqiao Wang",
"Yafeng Deng"
] | [] | http://arxiv.org/abs/2608.06243v2 | VERIFIED_LIVE | [
"https://github.com/DBtxy/DASH-OPSD"
] | [
200
] | [
true
] | Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models using automatically verifiable outcome signals, but these signals are typically sparse and at the sequence-level. On-policy self-distillation (OPSD) mitigates this sparsity by querying a privileged teacher... | [
0.029650000855326653,
-0.12095200270414352,
0.09997399896383286,
-0.007337000221014023,
0.08378200232982635,
-0.045680999755859375,
0.0402039997279644,
0.030590999871492386,
0.017230000346899033,
-0.012784999795258045,
-0.05911799892783165,
-0.03678499907255173,
0.012423000298440456,
0.075... | [
-0.11306899785995483,
-0.09363999962806702,
0.09060800075531006,
0.030717000365257263,
0.024637000635266304,
-0.03092999942600727,
-0.0583220012485981,
-0.008809000253677368,
0.084307000041008,
-0.03583399951457977,
-0.02522199973464012,
-0.0004939999780617654,
-0.006767000071704388,
0.071... | LLM Fine-Tuning, Quantization & Model Optimization | Reinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language models using automatically verifiable outcome signals, but these signals are typically sparse and at the sequence-level. | On-policy self-distillation (OPSD) mitigates this sparsity by querying a privileged teacher at student-visited prefixes and providing dense token-level distributional supervision. | Experiments on three mathematical reasoning benchmarks across three model scales show that DASH improves over our matched vanilla OPSD reruns on every benchmark at all three scales. | [
"Reinforcement Learning (RL)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Benchmarks (+1) | 2026-08-10T16:37:50.002716 |
2608.06161v1 | iARCS: Iterative Agentic RL for Controllable 3D Scene Generation | 2026-08-06T15:30:21Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Saugat Adhikari | 5 | [
"Saugat Adhikari",
"Ashok Prasad Neupane",
"Pramish Paudel",
"Ajad Chhatkuli",
"Danda Pani Paudel"
] | [] | http://arxiv.org/abs/2608.06161v1 | NOT_DETECTED | [] | [] | [] | Synthetic 3D scene generation is increasingly used as a data source for computer vision and embodied AI, but existing generators often optimize perceptual realism without reliably satisfying task-critical functional constraints. This mismatch limits the usefulness of synthetic data for downstream training, where access... | [
-0.04715299978852272,
-0.0810680016875267,
-0.041402000933885574,
-0.0323060005903244,
0.009535999968647957,
-0.0035099999513477087,
-0.021014999598264694,
-0.035753000527620316,
0.023814000189304352,
0.04087400063872337,
-0.016258999705314636,
-0.07524999976158142,
0.060081999748945236,
0... | [
-0.027751000598073006,
-0.08963999897241592,
0.005417000036686659,
0.02721799910068512,
0.02407499961555004,
0.04744800180196762,
-0.01673799939453602,
-0.08900099992752075,
-0.011880000121891499,
0.03369399905204773,
-0.05481800064444542,
-0.10772199928760529,
0.05581299960613251,
0.06032... | LLM Fine-Tuning, Quantization & Model Optimization | Synthetic 3D scene generation is increasingly used as a data source for computer vision and embodied AI, but existing generators often optimize perceptual realism without reliably satisfying task-critical functional constraints. | We present iARCS, an iterative agentic reinforcement learning framework that adapts a pretrained scene generator to natural-language task requirements. | We further show that data generated by iARCS improves a base generator, supporting its value as a practical synthetic data generation tool rather than only a controllable scene editing method. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)",
"Multi-Agent System"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:50.021266 |
2608.06107v1 | Kastor: An efficient fine-tuning strategy for generative emulation of PDE simulations | 2026-08-06T14:41:11Z | [
"cs.LG",
"math.NA"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Guillaume Couairon | 7 | [
"Guillaume Couairon",
"Alexis Jacq",
"Yu-Han Wu",
"Renu Singh",
"Yana Hasson",
"Quentin Berthet",
"Romuald Elie"
] | [] | http://arxiv.org/abs/2608.06107v1 | NOT_DETECTED | [] | [] | [] | Machine learning offers a promising avenue to accelerate physical simulations by replacing computationally expensive traditional Partial Differential Equation (PDE) solvers with fast, differentiable surrogate models. However, standard auto-regressive ML emulators often suffer from error accumulation over long horizons ... | [
-0.04107699915766716,
-0.06772500276565552,
0.09677399694919586,
-0.006095999851822853,
-0.07010699808597565,
-0.07946199923753738,
-0.10656700283288956,
-0.06818100064992905,
-0.020130999386310577,
-0.04394499957561493,
-0.023746000602841377,
-0.06642899662256241,
-0.01761999912559986,
-0... | [
-0.06960400193929672,
-0.08979599922895432,
0.15051400661468506,
0.11163199692964554,
0.01107100024819374,
-0.03367700055241585,
-0.14072099328041077,
-0.04075799882411957,
-0.03788100183010101,
-0.025885000824928284,
-0.08664499968290329,
-0.04311100021004677,
-0.02922300063073635,
-0.022... | LLM Fine-Tuning, Quantization & Model Optimization | Machine learning offers a promising avenue to accelerate physical simulations by replacing computationally expensive traditional Partial Differential Equation (PDE) solvers with fast, differentiable surrogate models. | However, standard auto-regressive ML emulators often suffer from error accumulation over long horizons and struggle to capture the stochasticity of complex physical systems. | Our model achieves a 42.9% average reduction in forecasting compared to our reference based on the Walrus finetuning methodology, and outperforms Walrus for 8 out of 10 datasets on variance-normalized RMSE (VRMSE). | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:50.031293 |
2608.06069v1 | Training-Free Token-Level Steering for LLM Personalized Co-Writing | 2026-08-06T14:13:02Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Wenhao Mao | 7 | [
"Wenhao Mao",
"Chengbin Hou",
"Weixiao Wang",
"Jialiang Zhu",
"Min Liu",
"Yibin Hao",
"Hairong Lv"
] | [] | http://arxiv.org/abs/2608.06069v1 | NOT_DETECTED | [] | [] | [] | While Large Language Models (LLMs) show great promise for personalization, they often lack specialized domain knowledge. Conventional solutions like fine-tuning struggle with high computational costs and rapid data updates, while Retrieval-Augmented Generation fails to provide fine-grained, token-level steering. Furthe... | [
-0.02185400016605854,
-0.08037000149488449,
0.04462200030684471,
0.009273000061511993,
0.05476300045847893,
0.02889000065624714,
0.08190800249576569,
0.041682999581098557,
-0.008205000311136246,
-0.0241870004683733,
-0.03865800052881241,
-0.01371300034224987,
0.05188500136137009,
-0.038617... | [
-0.029069999232888222,
-0.09503699839115143,
0.03996799886226654,
0.04342399910092354,
-0.002363000065088272,
0.0015439999988302588,
0.0405460000038147,
0.022796999663114548,
0.011723999865353107,
-0.004681999795138836,
-0.04101400077342987,
0.0282990001142025,
0.042226001620292664,
0.0070... | LLM Fine-Tuning, Quantization & Model Optimization | While Large Language Models (LLMs) show great promise for personalization, they often lack specialized domain knowledge. | To this end, we introduce SteerWrite, a training-free framework designed for personalized co-writing. | Experiments demonstrate that SteerWrite achieves state-of-the-art performance across diverse datasets, metrics, and models, significantly reducing human editing effort. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:50.036362 |
2608.05993v1 | Clinical Communication Processing with Models Trained on LLM-Generated Synthetic Data: A Structured Survey and Novel Application Case Studies | 2026-08-06T13:04:42Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Alexander Apartsin | 2 | [
"Alexander Apartsin",
"Yehudit Aperstein"
] | [] | http://arxiv.org/abs/2608.05993v1 | NOT_DETECTED | [] | [] | [] | Much clinical value is conveyed not through structured records but through communication: exchanges in which patients describe symptoms, clinicians reason and give instructions, ambulances hand over to emergency departments, and nurses pass on a shift. Such language differs from tabular data because meaning depends on ... | [
-0.010036000050604343,
-0.06992900371551514,
-0.005038000177592039,
-0.04465600103139877,
-0.027542000636458397,
-0.05554100126028061,
-0.09394600242376328,
0.041728999465703964,
0.0579649992287159,
-0.042594000697135925,
-0.04156100004911423,
-0.0020039998926222324,
0.019315000623464584,
... | [
0.016249999403953552,
-0.0007880000048317015,
0.02609900012612343,
-0.026763999834656715,
0.008197000250220299,
0.022105000913143158,
-0.019228000193834305,
0.08450599759817123,
0.07169000059366226,
-0.06430000066757202,
-0.05289300158619881,
0.025471000000834465,
0.013237999752163887,
0.0... | LLM Fine-Tuning, Quantization & Model Optimization | Much clinical value is conveyed not through structured records but through communication: exchanges in which patients describe symptoms, clinicians reason and give instructions, ambulances hand over to emergency departments, and nurses pass on a shift. | Such language differs from tabular data because meaning depends on speaker role, intent, causality, uncertainty, omission, and channel noise. | The main limitation is that most studies evaluate on held-out synthetic communication, while train-on-synthetic, test-on-authentic evidence remains limited. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:50.045110 |
2608.05813v1 | Cautious Context Steering for Language Model Personalization | 2026-08-06T09:45:25Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Gihoon Kim | 7 | [
"Gihoon Kim",
"Jeyoung Lee",
"Suhan Woo",
"Sekwon Oh",
"Minsu Jeon",
"Hyounsoo Han",
"Euntai Kim"
] | [] | http://arxiv.org/abs/2608.05813v1 | NOT_DETECTED | [] | [] | [] | Personalizing language models (LMs) to individual user preferences is essential for aligning responses with diverse goals and backgrounds. Existing methods typically train a separate adapter for each user or learn a reward model whose scores depend on the user. Despite explicitly optimizing for each user, these methods... | [
0.015659000724554062,
-0.028402000665664673,
-0.013381999917328358,
-0.013636000454425812,
0.08826100081205368,
0.020224999636411667,
0.21063800156116486,
0.019404999911785126,
-0.024764999747276306,
-0.0789790004491806,
0.0014609999489039183,
-0.0836540013551712,
0.14377500116825104,
0.00... | [
-0.004220000002533197,
-0.05582199990749359,
-0.017842000350356102,
0.03295300155878067,
0.10057699680328369,
0.014798000454902649,
0.12216299772262573,
0.031231999397277832,
-0.014055999927222729,
-0.051670998334884644,
-0.014821000397205353,
-0.10496199876070023,
0.10529199987649918,
0.0... | LLM Fine-Tuning, Quantization & Model Optimization | Personalizing language models (LMs) to individual user preferences is essential for aligning responses with diverse goals and backgrounds. | Existing methods typically train a separate adapter for each user or learn a reward model whose scores depend on the user. | A single CCS adapter trained on only one dataset improves generation quality both in-domain and across four out-of-distribution personalization benchmarks, demonstrating robust generalization to new users and domains. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:50.051922 |
2608.05802v1 | On-Policy Delta Distillation for Multilingual Math Reasoning | 2026-08-06T09:37:49Z | [
"cs.CL",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Byeongho Heo | 4 | [
"Byeongho Heo",
"Jaehui Hwang",
"Sangdoo Yun",
"Dongyoon Han"
] | [] | http://arxiv.org/abs/2608.05802v1 | NOT_DETECTED | [] | [] | [] | On-Policy Distillation (OPD) is emerging as a promising alternative to reinforcement learning for LLM post-training, yet its effectiveness in multilingual settings remains underexplored. We study OPD and its advanced variant, On-Policy Delta Distillation (OPD$^2$), for mathematical reasoning in English, Korean, and Jap... | [
-0.03838000074028969,
-0.04830799996852875,
0.08820799738168716,
-0.04723700135946274,
-0.0479779988527298,
-0.061014000326395035,
0.05897799879312515,
0.007908999919891357,
0.034074001014232635,
0.0042849997989833355,
-0.024302000179886818,
-0.031254999339580536,
-0.0026279999874532223,
0... | [
-0.0644029974937439,
-0.050916001200675964,
0.13612100481987,
0.013202999718487263,
-0.016397999599575996,
-0.021863000467419624,
0.03915699943900108,
-0.034147001802921295,
0.07217899709939957,
-0.012942000292241573,
0.002876000013202429,
0.041207000613212585,
-0.0027379998937249184,
0.00... | LLM Fine-Tuning, Quantization & Model Optimization | On-Policy Distillation (OPD) is emerging as a promising alternative to reinforcement learning for LLM post-training, yet its effectiveness in multilingual settings remains underexplored. | We study OPD and its advanced variant, On-Policy Delta Distillation (OPD$^2$), for mathematical reasoning in English, Korean, and Japanese. | Experiments with Qwen3 show that OPD$^2$ consistently outperforms the original OPD, with particularly strong improvements in Korean and Japanese, and generally narrows the English-Korean performance gap. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:50.056938 |
2608.05734v1 | Subliminal Learning is Non-Semantic Distillation | 2026-08-06T08:18:13Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Ethan Hadley | 2 | [
"Ethan Hadley",
"Eren Gultepe"
] | [] | http://arxiv.org/abs/2608.05734v1 | NOT_DETECTED | [] | [] | [] | Subliminal Learning (SL) is a surprising type of generalization displayed by modern language models. It allows the transfer of a bias or behavior from a teacher model to a student by distilling from seemingly unrelated or random synthetic data from the teacher. This presents challenges in ensuring AI systems remain pre... | [
0.08789200335741043,
-0.12268699705600739,
0.09493300318717957,
-0.013434999622404575,
0.05745600163936615,
-0.052838001400232315,
0.04381600022315979,
-0.012498999945819378,
0.07564699649810791,
-0.04152199998497963,
-0.007577000185847282,
-0.0747160017490387,
-0.018386000767350197,
0.017... | [
-0.00027600000612437725,
-0.10114099830389023,
0.04246800020337105,
0.009286999702453613,
0.09375900030136108,
0.008899999782443047,
0.06218799948692322,
-0.057436998933553696,
0.0839499980211258,
0.012257999740540981,
0.014670000411570072,
-0.049525998532772064,
0.04073899984359741,
0.023... | LLM Fine-Tuning, Quantization & Model Optimization | Subliminal Learning (SL) is a surprising type of generalization displayed by modern language models. | This presents challenges in ensuring AI systems remain predictable and are trained safely, as standard auditing of the input data would not catch the hidden subliminal signal. | Analysis of the activations of the student models that have been trained on steered and prompted data demonstrates that students inherit not just the semantic meaning of the teacher's bias, but also the type of intervention that was used to apply it: steered students imitate steering vectors, prompted students do not. | [
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:50.067079 |
2608.05724v1 | Sparse Mutual Information Graph Averaging for Improving Random Indexing Embeddings | 2026-08-06T08:07:18Z | [
"cs.CL",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Sriram Loganathan | 5 | [
"Sriram Loganathan",
"Gokul Anand",
"Aung Bo Bo",
"Yourui Shao",
"William B. Andreopoulos"
] | [] | http://arxiv.org/abs/2608.05724v1 | NOT_DETECTED | [] | [] | [] | Sparse word embedding pipelines can avoid dense co-occurrence matrix materialization, dense factorization, and gradient training while still relying on sparse global corpus statistics. This paper studies Random Indexing (RI) vectors refined by weighted averaging on a sparse Positive Pointwise Mutual Information (PPMI) ... | [
0.02802100032567978,
-0.06419999897480011,
0.06803199648857117,
0.02706499956548214,
0.06300999969244003,
0.038015998899936676,
0.035071998834609985,
0.005609000101685524,
0.02774200029671192,
-0.04427900165319443,
0.025126999244093895,
0.046052999794483185,
0.07683499902486801,
0.07796400... | [
-0.007453000172972679,
-0.08698700368404388,
0.049754999577999115,
0.009271999821066856,
0.03439300134778023,
0.006095000077039003,
-0.04254699870944023,
0.0064489999786019325,
0.0067170001566410065,
-0.017148999497294426,
0.02131900005042553,
0.017847999930381775,
0.035287998616695404,
0.... | LLM Fine-Tuning, Quantization & Model Optimization | Sparse word embedding pipelines can avoid dense co-occurrence matrix materialization, dense factorization, and gradient training while still relying on sparse global corpus statistics. | This paper studies Random Indexing (RI) vectors refined by weighted averaging on a sparse Positive Pointwise Mutual Information (PPMI) graph. | On the fairytales dataset, PPMI top-K=50 graph averaging improves RI with accuracy going from 19.4+-0.7% to 30.7+-2.9%, and performing best with a seed42 of 34.6%. | [
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:50.080131 |
2608.05705v1 | Spectral Aliasing Pretext: A novel task for Self-Supervised fault diagnosis in rotating machinery | 2026-08-06T07:46:51Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Victor Gialis | 4 | [
"Victor Gialis",
"Maxime Metz",
"David Esteve",
"Abdenour Soualhi"
] | [] | http://arxiv.org/abs/2608.05705v1 | NOT_DETECTED | [] | [] | [] | Deep learning is a new way for machinery fault diagnosis but requires extensive labeled data, a scarce resource in industrial settings. We propose Spectral Aliasing Pretext (SAP), a self-supervised learning method that pretrains models on unlabeled vibration data by exploiting spectral aliasing. We deliberately undersa... | [
-0.031615000218153,
-0.05043099820613861,
0.08567599952220917,
0.008345999754965305,
0.010591999627649784,
-0.021587999537587166,
0.05081300064921379,
0.019905000925064087,
-0.03705200180411339,
-0.0031969998963177204,
0.023310000076889992,
-0.04048600047826767,
0.03276899829506874,
-0.008... | [
-0.04919600114226341,
-0.09018299728631973,
0.09556300193071365,
0.06110500171780586,
0.01144499983638525,
0.0033110000658780336,
0.04744900017976761,
0.007734000217169523,
-0.07317599654197693,
-0.04473099857568741,
-0.008950999937951565,
-0.04770699888467789,
-0.06682299822568893,
0.0507... | LLM Fine-Tuning, Quantization & Model Optimization | Deep learning is a new way for machinery fault diagnosis but requires extensive labeled data, a scarce resource in industrial settings. | We propose Spectral Aliasing Pretext (SAP), a self-supervised learning method that pretrains models on unlabeled vibration data by exploiting spectral aliasing. | In contrast, full fine-tuning, including fully supervised training, does not lead to more stable or better results. | [
"Large Language Model (LLM)",
"Transformer Architecture"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:50.090664 |
2608.05587v1 | StepReflect: Structured UI Transition Reflection for Mobile GUI Agents | 2026-08-06T04:17:27Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Linqiang Guo | 6 | [
"Linqiang Guo",
"Wei Liu",
"Li Gu",
"Yang Wang",
"Tse-Hsun",
"Chen"
] | [
"Peter"
] | http://arxiv.org/abs/2608.05587v1 | NOT_DETECTED | [] | [] | [] | Autonomous mobile GUI agents require accurate action reflection for reliable long-horizon execution. Existing approaches rely on open-ended multimodal reasoning after each action, which is costly and poorly matched to the structured nature of GUI state transitions. We propose StepReflect, which formulates per-step GUI ... | [
-0.06482499837875366,
0.007652000058442354,
0.05324900150299072,
-0.05990200117230415,
0.02679000049829483,
-0.05169599875807762,
-0.0004130000015720725,
0.01640300080180168,
0.0010999999940395355,
-0.05150999873876572,
-0.027778999879956245,
-0.07750599831342697,
0.02995700016617775,
0.00... | [
-0.04220600053668022,
-0.07285299897193909,
0.06849599629640579,
-0.0573900006711483,
0.08216799795627594,
-0.06589700281620026,
-0.03268799930810928,
0.02184700034558773,
-0.006463999859988689,
-0.007120000198483467,
-0.04851299896836281,
-0.08158999681472778,
0.029180999845266342,
0.0611... | LLM Fine-Tuning, Quantization & Model Optimization | Autonomous mobile GUI agents require accurate action reflection for reliable long-horizon execution. | We propose StepReflect, which formulates per-step GUI reflection as supervised structured prediction conditioned on explicit transition specifications and paired visual evidence. | These results establish StepReflect as a practical, locally deployable alternative to repeated frontier-model reflection for long-horizon mobile GUI agents. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Knowledge Distillation"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:50.100767 |
2608.05541v1 | Hyper-ES: Effective Evolution Strategies for LLM Reasoning via Descent Direction Merging | 2026-08-06T02:39:10Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yu Gu | 6 | [
"Yu Gu",
"Zhi Zheng",
"Yunpeng Ba",
"Xialiang Tong",
"Mingxuan Yuan",
"Zhenkun Wang"
] | [] | http://arxiv.org/abs/2608.05541v1 | VERIFIED_LIVE | [
"https://github.com/kuangrepi/Hyper-ES"
] | [
200
] | [
true
] | Evolution Strategy (ES) is a promising alternative to gradient-based fine-tuning for resource-constrained Large Language Model (LLM) reasoning. However, directly applying ES to billion-parameter LLMs is highly ineffective. In such high-dimensional parameter spaces, most random perturbations are nearly orthogonal to use... | [
0.013633999973535538,
-0.03997199982404709,
0.0343099981546402,
-0.03637700155377388,
0.08414699882268906,
-0.08403699845075607,
-0.012306000106036663,
-0.013460000045597553,
-0.05195700004696846,
0.027338000014424324,
-0.021477999165654182,
-0.06292200088500977,
0.003068000078201294,
0.02... | [
-0.02990099973976612,
-0.10027100145816803,
0.06539499759674072,
0.025874000042676926,
0.06536000221967697,
-0.008084000088274479,
-0.0071299998089671135,
-0.03132700175046921,
-0.013303999789059162,
-0.01964999921619892,
-0.034797001630067825,
-0.01524099987000227,
0.004273999948054552,
-... | LLM Fine-Tuning, Quantization & Model Optimization | Evolution Strategy (ES) is a promising alternative to gradient-based fine-tuning for resource-constrained Large Language Model (LLM) reasoning. | We propose Hyper-ES, a subspace-based ES framework that avoids the weakness of ES in full-parameter search while exploiting its strength in low-dimensional optimization. | Results show that Hyper-ES consistently outperforms GRPO-LoRA by 1% while requiring 10% fewer space-consuming gradient updates. | [
"Large Language Model (LLM)",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 9 | Production-Ready (TRL 7-8) | Score 9/10. Live repo (+3); Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:51.088245 |
2608.05499v1 | APQF: Agentic Profiling-Guided Structured Pruning and Mixed-Precision Quantization with Adaptive Fine-Tuning | 2026-08-06T01:09:50Z | [
"cs.CV",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Sadegh Jafari | 5 | [
"Sadegh Jafari",
"Mohiuddin Bilwal",
"Fan Zhou",
"Brian Gelder",
"Ali Jannesari"
] | [] | http://arxiv.org/abs/2608.05499v1 | NOT_DETECTED | [] | [] | [] | Modern deep neural networks achieve strong performance, but their scale makes them costly and slow, especially on resource-constrained edge devices. Pruning and quantization address this, but rely on manual, expert choices and on algorithms that are hard to apply across architectures. Uniform settings also ignore how d... | [
0.04707400128245354,
-0.003370000049471855,
-0.021560000255703926,
-0.06290300190448761,
-0.0357309989631176,
-0.0015320000238716602,
0.00785799976438284,
-0.09481199830770493,
-0.021190999075770378,
-0.034940000623464584,
0.030629999935626984,
-0.012109000235795975,
0.004197999835014343,
... | [
-0.03257500007748604,
0.022633999586105347,
-0.016311999410390854,
0.01411799993366003,
0.018084999173879623,
-0.01846200041472912,
-0.057732999324798584,
-0.045455001294612885,
-0.015155999921262264,
-0.04839399829506874,
-0.051805999130010605,
-0.01622699946165085,
-0.022468000650405884,
... | LLM Fine-Tuning, Quantization & Model Optimization | Modern deep neural networks achieve strong performance, but their scale makes them costly and slow, especially on resource-constrained edge devices. | Pruning and quantization address this, but rely on manual, expert choices and on algorithms that are hard to apply across architectures. | Ablations show that uniform compression loses the most accuracy at matched compute, and that withholding profiling data from the planner hurts every model. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Multi-Agent System",
"Transformer Architecture",
"Quantization"
] | [
"18x reduction"
] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:51.129023 |
2608.05475v1 | KV-Skill: Forging Expertise in the Model's Native Language | 2026-08-05T23:46:56Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Zhaowei Han | 6 | [
"Zhaowei Han",
"Xiang Zhang",
"Bing Han",
"Kai Liu",
"Danqi Hu",
"Jie Liu"
] | [] | http://arxiv.org/abs/2608.05475v1 | VERIFIED_LIVE | [
"https://github.com/shawnzhg/KV-Skill"
] | [
200
] | [
true
] | Task knowledge is commonly stored either as text in the prompt or as an update to model weights. Text is modular but must be interpreted on every use, while weight adaptation makes the resulting capability difficult to load, remove, or share independently. We introduce KV-Skill, a design space of external factorized op... | [
0.03358199819922447,
-0.03681600093841553,
0.017348000779747963,
0.017954999580979347,
-0.04828999936580658,
0.010923000052571297,
0.01738700084388256,
0.06601600348949432,
-0.002506999997422099,
-0.011262999847531319,
-0.006467000115662813,
-0.07488100230693817,
0.021755000576376915,
0.02... | [
-0.04168400168418884,
-0.03477099910378456,
-0.015184000134468079,
0.06092999875545502,
-0.06513699889183044,
0.06401000171899796,
-0.0069599999114871025,
0.011649999767541885,
-0.012614999897778034,
0.0028609998989850283,
0.0034910000395029783,
-0.040787000209093094,
0.05901600047945976,
... | LLM Fine-Tuning, Quantization & Model Optimization | Task knowledge is commonly stored either as text in the prompt or as an update to model weights. | We introduce KV-Skill, a design space of external factorized operators that a frozen language model reads through a lightweight interface. | These results show that task knowledge can be acquired from text or experience, compressed into an external operator, and deployed separately from the backbone. | [
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 9 | Production-Ready (TRL 7-8) | Score 9/10. Live repo (+3); Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:52.090943 |
2608.05446v1 | EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents | 2026-08-05T22:29:20Z | [
"cs.LG",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Xuying Ning | 16 | [
"Xuying Ning",
"Dongqi Fu",
"Tianxin Wei",
"Hanqing Zeng",
"Yuanchen Bei",
"Bingxuan Li",
"Zihao Li",
"Qifan Wang",
"Xiang Shen",
"Yifan Wu",
"Jiayi Liu",
"Hong Li",
"Yinglong Xia",
"Xiangjun Fan",
"Hanghang Tong",
"Jingrui He"
] | [] | http://arxiv.org/abs/2608.05446v1 | NOT_DETECTED | [] | [] | [] | Long-horizon LLM agents increasingly rely on external execution support to maintain state, track progress, invoke tools, verify outcomes, and reuse experience across interactions. However, effective harness use raises two coupled challenges: state formation from noisy interaction traces and runtime control over externa... | [
0.020229000598192215,
-0.09622900187969208,
0.009701999835669994,
0.04015300050377846,
0.03799600154161453,
0.0022150001022964716,
0.0013259999686852098,
-0.006527000106871128,
-0.018293999135494232,
-0.04905200004577637,
0.029541000723838806,
-0.10078299790620804,
-0.02562600001692772,
0.... | [
0.002322999993339181,
-0.12224400043487549,
-0.022614000365138054,
0.049281999468803406,
0.021855000406503677,
-0.019386999309062958,
0.008747999556362629,
-0.04647599905729294,
-0.012943999841809273,
0.005013999994844198,
-0.025723999366164207,
-0.016714999452233315,
0.022957999259233475,
... | LLM Fine-Tuning, Quantization & Model Optimization | Long-horizon LLM agents increasingly rely on external execution support to maintain state, track progress, invoke tools, verify outcomes, and reuse experience across interactions. | We introduce EvoHarness-RL, which exposes Belief, Progress, and Experience (BPE) as policy-facing harness state. | These results suggest that long-horizon agents benefit from trainable policies for constructing and coordinating with external harness workspaces, beyond simply adding stronger tools or larger memories. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:52.135312 |
2608.05391v1 | Adaptive Arena-based Contestable Argumentative Network-of-Experts for Open-Ended Care Plan Coordination | 2026-08-05T20:18:45Z | [
"cs.AI",
"cs.MA"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Truong Thanh Hung Nguyen | 6 | [
"Truong Thanh Hung Nguyen",
"Hoang-Loc Cao",
"Phuc Ho",
"Phuc Truong Loc Nguyen",
"René Richard",
"Hung Cao"
] | [] | http://arxiv.org/abs/2608.05391v1 | NOT_DETECTED | [] | [] | [] | Care plan coordination demands synthesizing heterogeneous clinical, functional, and psychosocial information across multiple professional disciplines, where monolithic LLM pipelines cannot perform in a transparent or safe manner. We present CANOE (Contestable Argumentative Network-of-Experts), a multi-agent neuro-symbo... | [
0.055337999016046524,
0.007114000152796507,
-0.044891998171806335,
0.01627499982714653,
-0.007933000102639198,
0.022901000455021858,
0.007089000195264816,
0.050996001809835434,
0.055691998451948166,
-0.006233000196516514,
-0.10202000290155411,
0.02738099917769432,
0.0017770000267773867,
0.... | [
0.03780600056052208,
0.014901000075042248,
-0.0485600009560585,
-0.028077999129891396,
-0.03949600085616112,
-0.0032679999712854624,
-0.012137000449001789,
0.04032199829816818,
0.06643299758434296,
-0.01040900032967329,
-0.13377200067043304,
-0.009019999764859676,
-0.009778999723494053,
0.... | LLM Fine-Tuning, Quantization & Model Optimization | Care plan coordination demands synthesizing heterogeneous clinical, functional, and psychosocial information across multiple professional disciplines, where monolithic LLM pipelines cannot perform in a transparent or safe manner. | We present CANOE (Contestable Argumentative Network-of-Experts), a multi-agent neuro-symbolic framework that addresses these limitations through five modules: complexity assessment, adaptive team recruitment, role-based argumentative computation via an Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF), ... | and MedicalRAG using ROUGE-L, AlignScore, MEDCON F1, FKGL, and LLM-as-a-judge shows that medically fine-tuned models achieve the strongest clinical correctness and safety, while CANOE's argumentative structure provides faithful explanation and human contestability. | [
"Large Language Model (LLM)",
"Multi-Agent System"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:52.152271 |
2608.05326v1 | QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding | 2026-08-05T18:29:47Z | [
"cs.LG",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Ayushman Garg | 6 | [
"Ayushman Garg",
"Akshita Gupta",
"Shaswata Bhattacharya",
"Abhishek Gupta",
"Sandeep Kumar",
"Manoj Kumar"
] | [] | http://arxiv.org/abs/2608.05326v1 | NOT_DETECTED | [] | [] | [] | Autoregressive large language model inference is increasingly constrained by the memory footprint of the Key-Value (KV) cache. A dominant line of work reduces this footprint by evicting tokens that appear unimportant under attention-derived scores. However, such policies make an implicit irreversible decision: once a t... | [
0.009119999594986439,
-0.004898000042885542,
0.0252080000936985,
0.017424000427126884,
0.0815889984369278,
0.049789998680353165,
0.018379999324679375,
-0.06275200098752975,
0.09734000265598297,
-0.02511500008404255,
0.0939290001988411,
-0.03888000175356865,
-0.04062899947166443,
0.02170599... | [
0.005526999942958355,
-0.04820000007748604,
0.0038399999029934406,
0.07847700268030167,
0.04865099862217903,
0.030246999114751816,
0.023343000560998917,
-0.002481000032275915,
0.09424299746751785,
0.014853999949991703,
-0.021035000681877136,
0.004569999873638153,
-0.01631000079214573,
0.00... | LLM Fine-Tuning, Quantization & Model Optimization | Autoregressive large language model inference is increasingly constrained by the memory footprint of the Key-Value (KV) cache. | A dominant line of work reduces this footprint by evicting tokens that appear unimportant under attention-derived scores. | Across long-context understanding, retrieval, and reasoning benchmarks, QEvict consistently improves over representative eviction and quantization baselines, reducing missed attention and improving information retention | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Quantization"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:52.159482 |
2608.06417v1 | Latent Fact-Checking: Detecting Misinformation through Activation Engineering | 2026-08-05T18:00:14Z | [
"cs.LG",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Pedro Barcelos | 6 | [
"Pedro Barcelos",
"Otávio Parraga",
"Marcelo M. Mussi",
"Lucas M. Fraga",
"Lucas S. Kupssinskü",
"Rodrigo C. Barros"
] | [] | http://arxiv.org/abs/2608.06417v1 | VERIFIED_LIVE | [
"https://github.com/Malta-Lab/LaFaCt"
] | [
200
] | [
true
] | The proliferation of misinformation online has driven demand for scalable detection systems. While most existing approaches rely on surface-level linguistic features or external knowledge retrieval, we examine truthfulness as a geometric property of a language model's representation space. We introduce a misinformation... | [
0.007290000095963478,
0.002831999910995364,
0.003920999821275473,
0.054992999881505966,
0.05458800122141838,
-0.03403199836611748,
0.06363499909639359,
-0.05465000122785568,
0.03996000066399574,
0.014390000142157078,
0.007550999987870455,
-0.12012899667024612,
0.068572998046875,
0.00992400... | [
-0.0381149984896183,
-0.07434699684381485,
0.01006300002336502,
0.05355999991297722,
0.05864899978041649,
0.015115000307559967,
-0.0377810001373291,
0.02879600040614605,
0.11407999694347382,
-0.05595099925994873,
-0.028824999928474426,
-0.10368800163269043,
0.05484599992632866,
0.088074997... | LLM Fine-Tuning, Quantization & Model Optimization | The proliferation of misinformation online has driven demand for scalable detection systems. | While most existing approaches rely on surface-level linguistic features or external knowledge retrieval, we examine truthfulness as a geometric property of a language model's representation space. | We evaluate the method across 11 models from the Gemma, Llama, and Qwen families, ranging from 270M to 12B parameters, on three fact-checking benchmarks: AVeriTeC, LIAR, and FACTors. | [
"Large Language Model (LLM)",
"Transformer Architecture"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Benchmarks (+1) | 2026-08-10T16:37:52.997089 |
2608.05104v1 | BnBERT-iPET: Sparse Few-Shot Language Modeling for Bengali via Lottery Ticket Pruning | 2026-08-05T17:42:33Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Sajib Hossain | 5 | [
"Sajib Hossain",
"Md Kamrus Samad",
"Anan Ghosh",
"Labib Imam Chowdhury",
"Nabeel Mohammed"
] | [] | http://arxiv.org/abs/2608.05104v1 | NOT_DETECTED | [] | [] | [] | Deep neural networks have shown impressive success in NLP tasks owing to their complex structure and huge number of edges. Achieving state-of-the-art performance in natural language processing with a large pre-trained model such as BERT is expensive and time-consuming, carries a large carbon footprint, and is difficult... | [
0.006697000004351139,
0.029704000800848007,
0.0012809999752789736,
-0.04771500080823898,
-0.006014000158756971,
0.11859499663114548,
0.037985000759363174,
-0.013267000205814838,
0.04986799880862236,
-0.034612998366355896,
0.011098000220954418,
-0.04222499951720238,
0.04519899934530258,
0.0... | [
-0.08788300305604935,
-0.014514000155031681,
0.026480000466108322,
0.05322299897670746,
-0.019985999912023544,
0.0837090015411377,
-0.017092999070882797,
-0.028957000002264977,
0.005442999769002199,
-0.06637699902057648,
-0.05449699983000755,
-0.0617309994995594,
-0.026903999969363213,
0.0... | LLM Fine-Tuning, Quantization & Model Optimization | Deep neural networks have shown impressive success in NLP tasks owing to their complex structure and huge number of edges. | Achieving state-of-the-art performance in natural language processing with a large pre-trained model such as BERT is expensive and time-consuming, carries a large carbon footprint, and is difficult to realize on machines with minimal computational capability. | In this work, we introduce BnBERT-iPET, a sparse few-shot language modeling approach for Bengali, and experimentally show that a lightweight few-shot-learned language model retaining only 10% of the edges of an initial model such as BERT can perform neck and neck with much larger models on challenging tasks for a resou... | [
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:53.022234 |
2608.05076v1 | MultiPathFormer: Towards a Foundation Model for Multipath Wireless Propagation | 2026-08-05T17:20:16Z | [
"cs.LG",
"cs.AI",
"eess.SP"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Blessed Guda | 3 | [
"Blessed Guda",
"Kayley Sze",
"Carlee Joe-Wong"
] | [] | http://arxiv.org/abs/2608.05076v1 | NOT_DETECTED | [] | [] | [] | Recent advances in machine learning have enabled training of wireless foundation models, which aim to support tasks such as channel estimation, beam prediction, and localization based on wireless signals. Existing wireless foundation models typically pretrain on channel tensors using masked reconstruction over subcarri... | [
-0.011098000220954418,
-0.0979200005531311,
-0.01904899999499321,
-0.032680001109838486,
0.027493000030517578,
-0.07722199708223343,
-0.0717879980802536,
0.0327019989490509,
-0.07492800056934357,
-0.05933099985122681,
-0.01712699979543686,
0.009998000226914883,
0.043317001312971115,
0.0041... | [
-0.007978999987244606,
-0.17506499588489532,
0.0255730003118515,
0.07271800190210342,
0.055337000638246536,
-0.01510199997574091,
-0.061719998717308044,
-0.053252000361680984,
-0.08138100057840347,
-0.05859399959445,
-0.04217199981212616,
0.016232000663876534,
-0.01014500018209219,
0.06003... | LLM Fine-Tuning, Quantization & Model Optimization | Recent advances in machine learning have enabled training of wireless foundation models, which aim to support tasks such as channel estimation, beam prediction, and localization based on wireless signals. | In this work, we propose to instead use multipath propagation as the fundamental pretraining object. | These results show that path-level pretraining can learn reusable representations of wireless propagation. | [
"Large Language Model (LLM)",
"Transformer Architecture"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:53.037750 |
2608.05045v1 | Gradient Immunity: Null-Space Resistance to Malicious Fine-Tuning | 2026-08-05T16:55:59Z | [
"cs.CR",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yuxuan Huang | 4 | [
"Yuxuan Huang",
"Xingyu Zeng",
"Tianhang Zheng",
"Chaochao Lu"
] | [] | http://arxiv.org/abs/2608.05045v1 | VERIFIED_LIVE | [
"https://github.com/OpenCausaLab/Gradient-Immunity"
] | [
200
] | [
true
] | Released aligned large language models remain vulnerable to malicious downstream finetuning. Existing defenses are largely designed for the fine-tuning-as-a-service (FTaaS) paradigm or rely on downstream users to follow additional safety procedures, and therefore do not directly address the setting we study: a provider... | [
-0.07430599629878998,
-0.06518600136041641,
0.020774999633431435,
-0.035360001027584076,
0.034035999327898026,
0.029131999239325523,
0.040824998170137405,
-0.015157000161707401,
-0.034717001020908356,
0.004867999814450741,
0.014979000203311443,
-0.026511000469326973,
0.032412998378276825,
... | [
-0.08291500061750412,
-0.05212400108575821,
0.006777000147849321,
0.017590999603271484,
0.07868999987840652,
0.032687000930309296,
-0.007261999882757664,
-0.00673400005325675,
0.06908699870109558,
0.02161799930036068,
0.013338999822735786,
-0.019290000200271606,
0.027233000844717026,
-0.04... | LLM Fine-Tuning, Quantization & Model Optimization | Released aligned large language models remain vulnerable to malicious downstream finetuning. | Existing defenses are largely designed for the fine-tuning-as-a-service (FTaaS) paradigm or rely on downstream users to follow additional safety procedures, and therefore do not directly address the setting we study: a provider controlled partially protected open-weight (PPOW) release setting in which most weights rema... | These results suggest that release-time representation-space blocking can raise the cost of malicious downstream adaptation without requiring downstream cooperation. | [
"Large Language Model (LLM)",
"Transformer Architecture",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Benchmarks (+1) | 2026-08-10T16:37:53.873926 |
2608.05255v1 | An Emerging Retail Portfolio Management Application: Personalized, Tax-Aware Reinforcement Learning with Natural Language Goals | 2026-08-05T16:20:11Z | [
"cs.LG",
"cs.AI",
"cs.CR"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Ramin Pishehvar | 1 | [
"Ramin Pishehvar"
] | [] | http://arxiv.org/abs/2608.05255v1 | NOT_DETECTED | [] | [] | [] | Retail investors lack access to the kind of personalized, tax-aware portfolio management that institutional clients take for granted -- existing robo-advisors use static, rule-based allocation, and institutional-grade systems require account minimums and technology stacks unavailable to individual investors. We present... | [
-0.025800999253988266,
0.03177199885249138,
-0.004038000013679266,
-0.015881000086665154,
-0.0014909999445080757,
0.06380400061607361,
0.139833003282547,
-0.00444599986076355,
0.013926000334322453,
0.00785099994391203,
0.022099999710917473,
-0.04354500025510788,
-0.05725900083780289,
-0.04... | [
-0.06489299982786179,
-0.04738499969244003,
-0.08418300002813339,
0.003963000141084194,
-0.0056859999895095825,
-0.03110099956393242,
0.016457000747323036,
-0.01969599910080433,
-0.02896999940276146,
-0.007613999769091606,
-0.019690999761223793,
-0.022367000579833984,
0.04314899817109108,
... | LLM Fine-Tuning, Quantization & Model Optimization | Retail investors lack access to the kind of personalized, tax-aware portfolio management that institutional clients take for granted -- existing robo-advisors use static, rule-based allocation, and institutional-grade systems require account minimums and technology stacks unavailable to individual investors. | We present a fully built, integration-tested application that closes this gap: a FastAPI backend and web dashboard that let a user describe an investment goal in plain language (e.g. "I want steady growth but need to sell some shares next month for a down payment"), routes that goal to one of six investment mandates, a... | The system is functionally complete and integration-tested end-to-end against a live brokerage API (Alpaca, paper-trading mode), including multi-user authentication, a trust first preview-before-apply confirmation flow, daily email digests, and an auditable action-integrity chain, but has not yet been opened to real en... | [
"Reinforcement Learning (RL)",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:37:53.917981 |
2608.05245v1 | Search2Skill: Skill Distillation Beyond Knowledge Boundaries Via Rubric-Based Reinforcement Learning | 2026-08-05T15:09:07Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Muyang Ye | 13 | [
"Muyang Ye",
"Tian Lan",
"Feihu Jiang",
"Yongshi Ye",
"Wuyunsiqin",
"Bin Zhu",
"Qianghuai Jia",
"Zhao Xu",
"Weihua Luo",
"Ye Wang",
"Jinyang Zhang",
"Longyue Wang",
"Lingfeng Bao"
] | [] | http://arxiv.org/abs/2608.05245v1 | NOT_DETECTED | [] | [] | [] | Reusable skills, which encapsulate the procedural knowledge required to solve real-world professional tasks, offer LLM-based agents a path toward self-evolution in expert domains. Existing self-evolving skill methods construct skills internally from the model's parametric knowledge or trajectories, and are therefore bo... | [
-0.06133599951863289,
-0.05025700107216835,
0.06616900116205215,
0.04370500147342682,
-0.06930699944496155,
-0.035082001239061356,
-0.008704000152647495,
-0.0640840008854866,
-0.06821899861097336,
-0.0323840007185936,
-0.08212000131607056,
0.04575499892234802,
0.03891399875283241,
0.037636... | [
-0.0013040000339969993,
-0.07251299917697906,
-0.046870000660419464,
0.030479999259114265,
-0.012384000234305859,
0.008712000213563442,
-0.03129599988460541,
0.003252000082284212,
-0.0644569993019104,
-0.03298100084066391,
-0.034015998244285583,
0.025181999430060387,
0.025526000186800957,
... | LLM Fine-Tuning, Quantization & Model Optimization | Reusable skills, which encapsulate the procedural knowledge required to solve real-world professional tasks, offer LLM-based agents a path toward self-evolution in expert domains. | Existing self-evolving skill methods construct skills internally from the model's parametric knowledge or trajectories, and are therefore bounded by what the model already knows. | Further analyses show that the gains arise from skill abstraction rather than raw retrieved evidence, and that the acquired skills transfer across model scales. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:53.933007 |
2608.04934v1 | State2State: Environment-Derived Mid-Training for LLM Agents | 2026-08-05T15:02:41Z | [
"cs.CL",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Xuanyu Lei | 9 | [
"Xuanyu Lei",
"Yiqi Zhu",
"Chenliang Li",
"Kaiming Liu",
"Peng Li",
"Ming Yan",
"Jieping Ye",
"Ya-Qin Zhang",
"Yang Liu"
] | [] | http://arxiv.org/abs/2608.04934v1 | NOT_DETECTED | [] | [] | [] | Training LLM agents commonly relies on supervised fine-tuning from expert trajectories or online reinforcement learning over human-specified tasks with handcrafted verifiers. Though effective, both remain bottlenecked by externally specified tasks and supervision signals, limiting the scalability and diversity of agent... | [
0.025498999282717705,
-0.048002999275922775,
0.06399500370025635,
-0.03646299988031387,
0.04231899976730347,
-0.0019000000320374966,
-0.018866999074816704,
0.0025500000920146704,
-0.08629799634218216,
-0.012164000421762466,
-0.07763300091028214,
-0.06982400268316269,
0.05016599968075752,
0... | [
0.036219000816345215,
-0.06094399839639664,
0.03165600076317787,
0.019874999299645424,
0.06447300314903259,
-0.047658998519182205,
-0.01012600027024746,
-0.0455549992620945,
-0.057176001369953156,
-0.007954999804496765,
-0.06307700276374817,
-0.09511899948120117,
0.04911600053310394,
0.065... | LLM Fine-Tuning, Quantization & Model Optimization | Training LLM agents commonly relies on supervised fine-tuning from expert trajectories or online reinforcement learning over human-specified tasks with handcrafted verifiers. | We propose State2State, an environment-derived mid-training method that converts explored environment states into training objectives, challenging agents to reach a specified target state. | As initialization for downstream RL, it further improves final performance and learning efficiency, with promising evidence of cross-environment generalization. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:37:53.948202 |
2608.05242v2 | Disentangling 3D Modeling from Spatial Reasoning | 2026-08-05T14:32:48Z | [
"cs.LG",
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Haoze Sun | 4 | [
"Haoze Sun",
"Jiequan Cui",
"Qingshan Xu",
"Richang Hong"
] | [] | http://arxiv.org/abs/2608.05242v2 | NOT_DETECTED | [] | [] | [] | In this work, we explore an alternative paradigm for spatial reasoning by explicitly disentangling 3D perception from reasoning, rather than jointly acquiring implicit 3D perception and reasoning through large-scale training. Our key observation is that modern perception models excel at estimating continuous 3D geometr... | [
-0.04455899819731712,
-0.09524200111627579,
0.0547420009970665,
0.0009500000160187483,
-0.0019249999895691872,
-0.09318099915981293,
0.0165219996124506,
0.01501499954611063,
0.03310300037264824,
0.02573700062930584,
-0.04390899837017059,
-0.042817000299692154,
0.010427000001072884,
0.12597... | [
0.030534999445080757,
-0.11750999838113785,
0.06225699931383133,
0.023612000048160553,
0.036309998482465744,
-0.030083000659942627,
0.005663000047206879,
-0.006705999840050936,
0.036302000284194946,
0.01844700053334236,
-0.07169599831104279,
-0.037092000246047974,
-0.0005949999904260039,
0... | LLM Fine-Tuning, Quantization & Model Optimization | In this work, we explore an alternative paradigm for spatial reasoning by explicitly disentangling 3D perception from reasoning, rather than jointly acquiring implicit 3D perception and reasoning through large-scale training. | Motivated by these complementary strengths, we propose the Disentangled Spatial Reasoner (DiSR), a simple yet effective framework that reconstructs the physical world into structured 3D evidence using off-the-shelf expert perception models and fine-tunes an LLM with LoRA to perform reasoning solely over this explicit g... | Beyond its strong performance, DiSR offers improved interpretability, modularity, and computational efficiency, demonstrating that explicit separation of perception and reasoning is a scalable and effective alternative paradigm to end-to-end modeling for spatial intelligence. | [
"Large Language Model (LLM)",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:53.963092 |
2608.04872v2 | A-SR: Self-Evolving Agentic LLMs for Symbolic Regression via Hierarchical Coordination | 2026-08-05T14:01:10Z | [
"cs.CL",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Wenxiao Zhao | 12 | [
"Wenxiao Zhao",
"Dong Liu",
"Kaiyi Xu",
"Feng Liu",
"Zhen Zhao",
"Fei Ben",
"Shu Wang",
"Wenhao Li",
"Ying Nian Wu",
"Fenghua Ling",
"Haobo Li",
"Lei Bai"
] | [] | http://arxiv.org/abs/2608.04872v2 | NOT_DETECTED | [] | [] | [] | Symbolic regression aims to discover closed-form equations from data, but existing LLM-guided methods often rely on a unified proposal loop that compresses heterogeneous search failures into a scalar score and a single prompt. We propose A-SR, a self-evolving agentic framework that shifts the control unit from expressi... | [
0.027759000658988953,
-0.11197099834680557,
0.029133999720215797,
-0.007075000088661909,
0.09315700083971024,
0.0008110000053420663,
-0.012577000074088573,
-0.05077100172638893,
0.06996200233697891,
0.028579000383615494,
0.032485999166965485,
0.0035840000491589308,
0.061911001801490784,
0.... | [
-0.04339199885725975,
-0.05074400082230568,
-0.024082999676465988,
0.024963000789284706,
0.08364400267601013,
0.028024999424815178,
-0.08271200209856033,
0.02529500052332878,
0.015317999757826328,
0.049828000366687775,
-0.04847100004553795,
0.007189000025391579,
0.07068099826574326,
0.0627... | LLM Fine-Tuning, Quantization & Model Optimization | Symbolic regression aims to discover closed-form equations from data, but existing LLM-guided methods often rely on a unified proposal loop that compresses heterogeneous search failures into a scalar score and a single prompt. | We propose A-SR, a self-evolving agentic framework that shifts the control unit from expression edits to role-conditioned evidence views. | Averaged over the four LSR-Synth scientific domains in LLM-SRBench, A-SR improves Acc@0.01 over baselines from 25.79% to 48.30% with Llama3.1-8B, while A-SR-LoRA improves the corresponding Qwen3-4B result from 24.58% to 38.29%. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Multi-Agent System",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:53.978950 |
2608.04788v1 | Agentic Reinforcement Learning with Observation-Calibrated Self-Distillation | 2026-08-05T12:52:14Z | [
"cs.LG",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yi Yang | 11 | [
"Yi Yang",
"Cong Qin",
"Xiaodan Liu",
"Chishui Chen",
"Qing Dong",
"Yan Zhang",
"Cao Liu",
"Zhao Yang",
"Lu Pan",
"Jiaye Lin",
"Yi Feng"
] | [] | http://arxiv.org/abs/2608.04788v1 | VERIFIED_LIVE | [
"https://github.com/yiy1x/OCSD"
] | [
200
] | [
true
] | Large language model agents are commonly trained through reinforcement learning with sparse trajectory-level rewards, which offer limited guidance on how strongly individual tokens should be updated. On-Policy Self-Distillation (OPSD) addresses this by re-scoring generated tokens under a privileged replay view to obtai... | [
-0.025299999862909317,
-0.0851759985089302,
0.06063900142908096,
0.021755000576376915,
-0.026264000684022903,
0.0036150000523775816,
0.07428999990224838,
-0.05751900002360344,
0.043956998735666275,
0.0010029999539256096,
-0.01464099995791912,
-0.01660499908030033,
0.028008999302983284,
0.0... | [
-0.015250000171363354,
-0.1240679994225502,
0.0887330025434494,
0.010320999659597874,
0.07427199929952621,
0.028932999819517136,
0.008038000203669071,
-0.019005000591278076,
0.08489400148391724,
0.012254999950528145,
-0.05086499825119972,
-0.006459999829530716,
0.010262000374495983,
0.0094... | LLM Fine-Tuning, Quantization & Model Optimization | Large language model agents are commonly trained through reinforcement learning with sparse trajectory-level rewards, which offer limited guidance on how strongly individual tokens should be updated. | To resolve this confounding, we propose Observation-Calibrated Self-Distillation (OCSD), which contrasts two structurally matched replay views, Full and Observation-Ablated, differing only in whether the actual future observation is present, to derive an observation residual that discounts score changes shared by the r... | Experiments on ALFWorld, WebShop, and Search-QA across three Qwen3 model scales show that OCSD consistently outperforms strong baselines. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Multi-Agent System"
] | [] | Explosive (>50/mo) | 491 | 7 | Prototype (TRL 4-6) | Score 7/10. Live repo (+3) | 2026-08-10T16:37:54.870327 |
2608.04765v1 | Explicit Language Memory for Long-Horizon Planning in Vision-Language-Action Models | 2026-08-05T12:32:15Z | [
"cs.RO",
"cs.AI",
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Houze Xu | 3 | [
"Houze Xu",
"Jizhong Li",
"Ziyi Ye"
] | [] | http://arxiv.org/abs/2608.04765v1 | NOT_DETECTED | [] | [] | [] | Vision-language-action (VLA) models provide a unified paradigm for connecting visual perception, language understanding, and robotic control. However, existing VLA models still face major challenges in long-horizon tasks: sparse expert demonstrations constrain cross-task compositional generalization; the non-Markovian ... | [
0.07188300043344498,
-0.0779229998588562,
0.013855000026524067,
-0.006163999903947115,
0.03416400030255318,
0.11801499873399734,
-0.0030350000597536564,
0.00558099988847971,
0.03558000177145004,
-0.011880000121891499,
0.005014999769628048,
-0.02186799980700016,
0.008725999854505062,
0.0590... | [
0.07449600100517273,
-0.0863649994134903,
0.015987999737262726,
0.018583999946713448,
0.02057800069451332,
0.03254399821162224,
-0.03451700136065483,
-0.002070999937132001,
0.022760000079870224,
-0.015122000128030777,
-0.015699999406933784,
-0.04028400033712387,
0.024915000423789024,
0.061... | LLM Fine-Tuning, Quantization & Model Optimization | Vision-language-action (VLA) models provide a unified paradigm for connecting visual perception, language understanding, and robotic control. | However, existing VLA models still face major challenges in long-horizon tasks: sparse expert demonstrations constrain cross-task compositional generalization; the non-Markovian nature of long-horizon tasks makes it difficult for policies conditioned only on current observations to maintain temporal consistency; limite... | The results demonstrate that explicit language memory improves the success rate and robustness of VLA models on complex long-horizon tasks while providing an interpretable semantic account of the decision process. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:54.910388 |
2608.04697v1 | Traceable LLM-Generated Hazard Scenarios for Operational Safety Analysis of Aviation Systems Using ASRS Reports | 2026-08-05T11:07:17Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Cristian Mascia | 4 | [
"Cristian Mascia",
"Roberto Pietrantuono",
"Daniel Rodriguez",
"Stefano Russo"
] | [] | http://arxiv.org/abs/2608.04697v1 | NOT_DETECTED | [] | [] | [] | Operational hazard analysis of aviation system operations must consider interactions among weather, ATC actions, airspace constraints, aircraft operations, and human factors - distinct from the functional hazard assessment applied at the aircraft-system level. We present an AI-assisted approach that generates candidate... | [
0.022268999367952347,
-0.045049998909235,
-0.06843499839305878,
0.0032790000550448895,
0.10101799666881561,
-0.0024139999877661467,
-0.04502899944782257,
0.07600700110197067,
-0.026172999292612076,
0.03615500032901764,
0.003418999956920743,
-0.0059710000641644,
0.007187000010162592,
0.0079... | [
0.05454299971461296,
-0.018592000007629395,
-0.010724999941885471,
0.06100999936461449,
0.10551699995994568,
0.020894000306725502,
0.0016899999463930726,
0.060058001428842545,
0.04233599826693535,
0.02706499956548214,
-0.056102998554706573,
-0.062268998473882675,
-0.010002999566495419,
0.0... | LLM Fine-Tuning, Quantization & Model Optimization | Operational hazard analysis of aviation system operations must consider interactions among weather, ATC actions, airspace constraints, aircraft operations, and human factors - distinct from the functional hazard assessment applied at the aircraft-system level. | We present an AI-assisted approach that generates candidate hazard scenarios from NASA's Aviation Safety Reporting System (ASRS). | We evaluate multiple large language models, zero-shot versus few-shot prompting, and optional fine-tuning, measuring how prompting and model choice affect the validity and realism of the generated structures and narratives. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:54.926812 |
2608.04576v1 | Causal Evidence Extraction and Triangulation in Crisis Reports using Large Language Models: A ReliefWeb-based Study | 2026-08-05T08:07:15Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yuanjun Zhang | 2 | [
"Yuanjun Zhang",
"Mourad Oussalah"
] | [] | http://arxiv.org/abs/2608.04576v1 | NOT_DETECTED | [] | [] | [] | Humanitarian reports are long, noisy, and multi-topic, making it difficult to consolidate decision-relevant causal evidence. We present a ReliefWeb study (2000-2024) and a two-stage Large Language Model (LLM) pipeline that extracts structured intervention-outcome records with direction and strength attributes. Query-co... | [
-0.006819999776780605,
0.000307999987853691,
0.05531900003552437,
0.04258599877357483,
0.07023400068283081,
0.08337999880313873,
-0.04129499942064285,
0.0514569990336895,
0.036428000777959824,
0.025018999353051186,
0.014700000174343586,
-0.026378000155091286,
0.03534099832177162,
0.0232239... | [
-0.021076999604701996,
0.02213199995458126,
0.03759099915623665,
0.04434699937701225,
0.10894300043582916,
0.06808199733495712,
-0.04020300135016441,
0.022401999682188034,
-0.005173000041395426,
0.042897000908851624,
-0.02637699991464615,
0.004602000117301941,
0.06247299909591675,
0.056846... | LLM Fine-Tuning, Quantization & Model Optimization | Humanitarian reports are long, noisy, and multi-topic, making it difficult to consolidate decision-relevant causal evidence. | We present a ReliefWeb study (2000-2024) and a two-stage Large Language Model (LLM) pipeline that extracts structured intervention-outcome records with direction and strength attributes. | In an expert-annotated dataset of 100 reports, the best closed-source LLM achieved a weighted F1 score of 90.73% with strong cost-efficiency, while Llama-3.1-8B with supervised fine-tuning reached 94.15% weighted F1 score. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:54.943609 |
2608.05207v1 | When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters | 2026-08-05T07:48:35Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Fangxin Wang | 7 | [
"Fangxin Wang",
"Ziyi Zhang",
"Diyi Zhuang",
"Langzhou He",
"Shiyu Wang",
"Baichuan Mo",
"Philip S. Yu"
] | [] | http://arxiv.org/abs/2608.05207v1 | NOT_DETECTED | [] | [] | [] | Frozen pretrained forecasters often fail in structured, recurring ways that are costly to repair through fine-tuning. We study corrective feature discovery: mining interpretable features of a frozen forecaster's residual to drive a lightweight post-hoc corrector. Prior automated feature engineering models the data-gene... | [
-0.009736999869346619,
-0.003017999930307269,
0.07065299898386002,
0.016016000881791115,
-0.0016690000193193555,
0.05987099930644035,
-0.018161000683903694,
-0.06157999858260155,
-0.054499998688697815,
-0.018219999969005585,
-0.013411999680101871,
0.032951001077890396,
0.0016210000030696392,... | [
-0.06165900081396103,
-0.038343001157045364,
0.032315999269485474,
0.09833700209856033,
0.034922000020742416,
-0.05778300017118454,
-0.10069599747657776,
-0.01670899987220764,
-0.0623369999229908,
-0.04362000152468681,
-0.10702600330114365,
-0.032235000282526016,
-0.016207000240683556,
-0.... | LLM Fine-Tuning, Quantization & Model Optimization | Frozen pretrained forecasters often fail in structured, recurring ways that are costly to repair through fine-tuning. | We present CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning), which keeps the backbone frozen and mines its residual with two complementary generators: a compositional search over the raw input channels, and a large language model (LLM) that proposes named feature combinations, b... | Across six public datasets and six frozen backbones, CRAFTER surpasses every dedicated feature-engineering system at every feature budget, roughly doubling the improvement achieved by the corrector alone and reducing the error of the weakest backbones by up to 27%. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:54.962550 |
2608.04488v1 | Energy- and Memory-Efficient PEFT Methods for Personalized On-Device SLMs on Consumer GPUs | 2026-08-05T06:20:57Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Kuanysh Akhmetzhanov | 2 | [
"Kuanysh Akhmetzhanov",
"Jurn-Gyu Park"
] | [] | http://arxiv.org/abs/2608.04488v1 | NOT_DETECTED | [] | [] | [] | Despite rapid advances in large language models (LLMs), deploying and personalizing them on resource-constrained devices remains impractical due to high VRAM, time, and energy costs. Parameter-Efficient Fine-Tuning (PEFT) of Small Language Models (SLMs) offers a promising alternative, yet few studies compare PEFT metho... | [
-0.03584500029683113,
-0.011575000360608101,
0.023435000330209732,
0.0309820007532835,
0.02467699907720089,
-0.09953899681568146,
0.1231670007109642,
0.053268998861312866,
-0.02631000056862831,
-0.03348100185394287,
0.03273500129580498,
0.06302300095558167,
0.01832900010049343,
-0.04192499... | [
-0.006300999782979488,
-0.012497999705374241,
-0.025043999776244164,
-0.00039000000106170774,
0.011982999742031097,
-0.011380000039935112,
0.05681199952960014,
0.05142499879002571,
-0.007613999769091606,
-0.0015310000162571669,
0.00874600000679493,
-0.032618001103401184,
-0.00136500003281980... | LLM Fine-Tuning, Quantization & Model Optimization | Despite rapid advances in large language models (LLMs), deploying and personalizing them on resource-constrained devices remains impractical due to high VRAM, time, and energy costs. | Parameter-Efficient Fine-Tuning (PEFT) of Small Language Models (SLMs) offers a promising alternative, yet few studies compare PEFT methods across architectures using both general and personalization benchmarks while accounting for energy consumption. | These results show that compact SLMs paired with PEFT provide a practical, energy-aware path to personalized on-device deployment, with the optimal method set by the dominant constraint: LoRA+ for energy and QLoRA for memory. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Transformer Architecture",
"LoRA / PEFT",
"Quantization"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:54.981669 |
2608.04433v1 | MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages | 2026-08-05T04:22:27Z | [
"cs.CL",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Qiongqiong Wang | 21 | [
"Qiongqiong Wang",
"Ai Ti Aw",
"Nancy F. Chen",
"Ying Lay Chiu",
"Yang Ding",
"Yingxu He",
"Ridong Jiang",
"Zhuohan Liu",
"Yanfeng Lu",
"Yi Ma",
"Muhammad Huzaifah",
"Nabilah Binte Md Johan",
"Nattadaporn Lertcheva",
"Pham Minh Duc",
"Sailor Hardik Bhupendra",
"Siti Umairah Binte Moham... | [] | http://arxiv.org/abs/2608.04433v1 | NOT_DETECTED | [] | [] | [] | We present MERaLiON-GR, a speech gender recognition system that performs binary classification (female / male) on English and Southeast Asian (SEA) languages. The model finetunes MERaLiON-SpeechEncoder-2, a large conformer based transformer pre-trained on a broad speech corpus, and applies parameter efficient fine-tuni... | [
-0.06907600164413452,
0.004484999924898148,
0.030633000656962395,
-0.03880300000309944,
-0.021448999643325806,
0.006312999874353409,
0.07018499821424484,
-0.1131109967827797,
-0.02950800023972988,
-0.09731700271368027,
-0.013419999741017818,
-0.09761899709701538,
-0.031244000419974327,
0.0... | [
-0.05288099870085716,
-0.04457699880003929,
0.01946599967777729,
-0.019843999296426773,
0.012457000091671944,
0.0044639999978244305,
-0.004794000182300806,
-0.06985200196504593,
-0.03989500179886818,
-0.07513300329446793,
-0.025547999888658524,
-0.13480499386787415,
-0.05022599920630455,
0... | LLM Fine-Tuning, Quantization & Model Optimization | We present MERaLiON-GR, a speech gender recognition system that performs binary classification (female / male) on English and Southeast Asian (SEA) languages. | The model finetunes MERaLiON-SpeechEncoder-2, a large conformer based transformer pre-trained on a broad speech corpus, and applies parameter efficient fine-tuning via Low-Rank Adaptation (LoRA) to adapt the encoder to the gender recognition task, and appends a multi-scale ECAPA-TDNN down stream network with attention ... | The results underscore the value of dedicated speech models in achieving accurate paralinguistic understanding and strong cross-lingual generalization. | [
"Large Language Model (LLM)",
"Transformer Architecture",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:37:54.998231 |
2608.04322v1 | DataRx: Missingness-Aware Sampling for Safer Large Language Model Task-Specific Fine-Tuning | 2026-08-05T01:05:13Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Junbo Zhang | 4 | [
"Junbo Zhang",
"Qianli Zhou",
"Xinyang Deng",
"Wen Jiang"
] | [] | http://arxiv.org/abs/2608.04322v1 | NOT_DETECTED | [] | [] | [] | Task-specific fine-tuning can improve the performance of large language models (LLMs) on downstream tasks. However, our study reveals that task-specific fine-tuning can also weaken the safety guardrails of aligned LLMs. A widely adopted strategy for preserving safety during fine-tuning is to incorporate safety data. Al... | [
-0.01069399993866682,
-0.046427998691797256,
0.015738999471068382,
0.018379999324679375,
0.11167100071907043,
0.0236200001090765,
0.06234300136566162,
0.004701999947428703,
0.015859000384807587,
-0.027735000476241112,
-0.02639099955558777,
-0.10888200253248215,
0.05246900022029877,
-0.0383... | [
0.0016740000573918223,
-0.08365999907255173,
0.03601599857211113,
-0.02968600019812584,
0.10936100035905838,
0.014152999967336655,
0.029983999207615852,
0.0014659999869763851,
-0.004978000186383724,
0.010131999850273132,
-0.06814400106668472,
-0.061778999865055084,
0.10306300222873688,
-0.... | LLM Fine-Tuning, Quantization & Model Optimization | Task-specific fine-tuning can improve the performance of large language models (LLMs) on downstream tasks. | In this paper, we propose DataRx, a missingness-aware sampling method for selecting safety-critical examples. | The results show that, with only 1% additional safety samples from BeaverTails, DataRx reduces the average attack success rate of Llama3-8B-Instruct across seven downstream tasks from 59.23% under random sampling to 13.70%. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:37:55.017107 |
2608.04213v1 | Attention-Only White-Box Transformer via LeJEPA-Based Self-Supervised Pretraining | 2026-08-04T20:26:12Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yang Bai | 6 | [
"Yang Bai",
"Linyuan Wang",
"Haoyang Jiang",
"Nuolin Sun",
"Libin Hou",
"Bin Yan"
] | [] | http://arxiv.org/abs/2608.04213v1 | NOT_DETECTED | [] | [] | [] | Existing studies on self-supervised learning for white-box networks typically decouple the derivation of white-box networks via optimization algorithms from self-supervised learning paradigms. In this work, we instead revisit the two components from a joint perspective. The LeJEPA-based self-supervised framework assume... | [
0.018157999962568283,
0.00016700000560376793,
0.03953099995851517,
0.001193000003695488,
-0.01508800033479929,
0.03917299956083298,
-0.040366001427173615,
0.011195999570190907,
-0.03654100000858307,
-0.046167999505996704,
-0.009344999678432941,
-0.013988000340759754,
-0.0354899987578392,
0... | [
-0.008078999817371368,
0.02895200066268444,
0.02240299992263317,
0.009335000067949295,
0.030006999149918556,
0.010018999688327312,
-0.03233100101351738,
0.027468999847769737,
-0.006256999913603067,
-0.031098999083042145,
-0.0603180006146431,
0.030703000724315643,
0.003208999987691641,
0.04... | LLM Fine-Tuning, Quantization & Model Optimization | Existing studies on self-supervised learning for white-box networks typically decouple the derivation of white-box networks via optimization algorithms from self-supervised learning paradigms. | The LeJEPA-based self-supervised framework assumes an isotropic Gaussian distribution as the optimal embedding distribution for downstream tasks, which is conceptually equivalent to the expansion term $R(Z)$ in the sparse rate reduction objective guiding white-box Transformer optimization. | Our model achieves competitive performance while reducing the parameter count by roughly $31\%$. | [
"Small Language Model (SLM)",
"Transformer Architecture",
"Knowledge Distillation"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:55.035721 |
2608.04084v1 | SpecDrop: Parameter-Free Category-Conditioned Routing for Modular Specialization | 2026-08-04T18:00:01Z | [
"cs.LG",
"cs.CL",
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Boyao Wang | 2 | [
"Boyao Wang",
"Zhihan Lei"
] | [] | http://arxiv.org/abs/2608.04084v1 | VERIFIED_LIVE | [
"https://github.com/Beryex/SpecDrop"
] | [
200
] | [
true
] | Mixture-of-experts (MoE) networks pursue specialization through learned routers, gates, and load-balancing losses, yet at matched total-parameter budgets learned routers can underperform equal-weight No-Routing baselines. Is the bottleneck the routing algorithm, or the alignment between training-signal granularity and ... | [
-0.016124999150633812,
-0.004871000070124865,
-0.02057100087404251,
-0.00495699979364872,
0.015475999563932419,
-0.00901699997484684,
0.03144799917936325,
-0.04104600101709366,
-0.10272599756717682,
-0.02528800070285797,
0.014677999541163445,
-0.05660799890756607,
-0.028147000819444656,
-0... | [
-0.03831300139427185,
-0.06487099826335907,
0.014361999928951263,
0.05850699916481972,
0.06019899994134903,
-0.014185000211000443,
0.026770999655127525,
0.019974999129772186,
-0.005688000004738569,
-0.07537399977445602,
-0.031525999307632446,
-0.020320000126957893,
-0.03875799849629402,
-0... | LLM Fine-Tuning, Quantization & Model Optimization | Mixture-of-experts (MoE) networks pursue specialization through learned routers, gates, and load-balancing losses, yet at matched total-parameter budgets learned routers can underperform equal-weight No-Routing baselines. | Is the bottleneck the routing algorithm, or the alignment between training-signal granularity and the target categories? | On fuzzy partitions, where training units span multiple categories (SlimPajama-6B language modeling with a 30M Transformer; SuperNI instruction tuning over Llama-3.2-1B with LoRA), the routing mechanism reduces to the matched No-Routing controls within seed noise, the null our thesis predicts. | [
"Large Language Model (LLM)",
"Transformer Architecture",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Quantization tool (+1) | 2026-08-10T16:37:55.940523 |
2608.04010v1 | ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs | 2026-08-04T17:59:58Z | [
"cs.CV",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yang Yang | 8 | [
"Yang Yang",
"Qinyu Zhao",
"Mouxiang Chen",
"Xiaohui Li",
"Lixin Gu",
"Wenhai Wang",
"Hongjie Zhang",
"Wenwei Zhang"
] | [] | http://arxiv.org/abs/2608.04010v1 | VERIFIED_LIVE | [
"https://github.com/YangYangGirl/ParVL"
] | [
200
] | [
true
] | Existing scaling strategies for Multimodal Large Language Models (MLLMs) typically expand either model parameters or sequential inference computation, incurring substantial memory or latency overhead. More importantly, most existing methods fail to alter the rigid, fixed computation allocation between the Vision Transf... | [
0.0330669991672039,
-0.1180960014462471,
0.01396500039845705,
-0.038325998932123184,
-0.012800999917089939,
-0.12067099660634995,
-0.02519400045275688,
-0.009297999553382397,
-0.005654999986290932,
0.029782000929117203,
0.0017460000235587358,
-0.06987299770116806,
0.009851999580860138,
0.0... | [
0.00824199989438057,
-0.13443699479103088,
-0.017903000116348267,
-0.045848000794649124,
0.0346050001680851,
-0.012996000237762928,
-0.08537500351667404,
0.03877300024032593,
0.04273099824786186,
-0.0541829988360405,
-0.009014000184834003,
-0.08169800043106079,
-0.01549800019711256,
0.0509... | LLM Fine-Tuning, Quantization & Model Optimization | Existing scaling strategies for Multimodal Large Language Models (MLLMs) typically expand either model parameters or sequential inference computation, incurring substantial memory or latency overhead. | More importantly, most existing methods fail to alter the rigid, fixed computation allocation between the Vision Transformer and the Large Language Model components, limiting task-specific optimization. | ParVL improves overall multimodal performance over same-recipe single-branch baselines, and the best evaluated vision--language allocation varies across tasks. | [
"Large Language Model (LLM)",
"Transformer Architecture"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Benchmarks (+1) | 2026-08-10T16:37:56.861284 |
2608.03979v1 | Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent | 2026-08-04T17:45:16Z | [
"cs.CV",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Zhen Fang | 20 | [
"Zhen Fang",
"Yu Zeng",
"Wenxuan Huang",
"Yiming Zhao",
"Shiting Huang",
"Tianfei Ren",
"Qi Lu",
"Qingnan Ren",
"Qisheng Su",
"Lionel Z. Wang",
"Qingyu Yin",
"Shuang Chen",
"Zehui Chen",
"Lin Chen",
"Zhenfei Yin",
"Yao Hu",
"Shaohui Lin",
"Wanli Ouyang",
"Shaosheng Cao",
"Feng ... | [] | http://arxiv.org/abs/2608.03979v1 | VERIFIED_LIVE | [
"https://github.com/Osilly/Vision-DeepResearch"
] | [
200
] | [
true
] | We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. Preliminary evaluations reveal two critical bottlenecks in current models: (1) modality bias, where agents bypa... | [
-0.03285500034689903,
-0.12996000051498413,
0.021851999685168266,
-0.02513900026679039,
0.049598000943660736,
0.02371699921786785,
-0.1258379966020584,
-0.03720799833536148,
0.0166229996830225,
-0.13708500564098358,
-0.004141000099480152,
-0.0679979994893074,
-0.01759300008416176,
0.069577... | [
-0.024692999199032784,
-0.17244599759578705,
0.02419400028884411,
-0.01706399954855442,
0.11542899906635284,
0.02501700073480606,
-0.049665000289678574,
-0.03896699845790863,
-0.011988000012934208,
-0.0600150004029274,
-0.051295001059770584,
-0.06181500107049942,
-0.03937099874019623,
0.08... | LLM Fine-Tuning, Quantization & Model Optimization | We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. | To address these challenges, we propose Video-DR, featuring a decoupled perception-exploration pipeline with stage-wise tool unlocking that compels exhaustive cross-frame visual grounding prior to web retrieval. | The 30B-A3B variant achieves 59.3%, competitive with Claude-4.5-Sonnet and demonstrating the effectiveness of our training paradigm even at compact scale. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 9 | Production-Ready (TRL 7-8) | Score 9/10. Live repo (+3); Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:37:57.686797 |
2608.03952v1 | TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring | 2026-08-04T17:16:14Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Dongjie Yang | 6 | [
"Dongjie Yang",
"Siyan Lin",
"Leixian Shen",
"Rui Sheng",
"Huamin Qu",
"Zixin Chen"
] | [] | http://arxiv.org/abs/2608.03952v1 | NOT_DETECTED | [] | [] | [] | Large language models (LLMs) are increasingly used to provide conversational practice for English-as-a-second-language (ESL) learners. Effective ESL tutoring, however, requires more than fluent response generation: a tutor must select an appropriate pedagogical action based on learner behavior and dialogue context. Hum... | [
0.04686100035905838,
-0.04501400142908096,
-0.005030000116676092,
-0.015309000387787819,
-0.031488001346588135,
-0.0014130000490695238,
0.010445999912917614,
0.006597000174224377,
-0.0067010000348091125,
0.044829998165369034,
0.028488000854849815,
-0.011385000310838223,
0.040516000241041183,... | [
-0.011261999607086182,
-0.0678969994187355,
0.028307000175118446,
0.004813000094145536,
-0.008590999990701675,
-0.06166600063443184,
0.03719399869441986,
0.007753999903798103,
0.015344000421464443,
-0.015664000064134598,
-0.000678999989759177,
-0.04400299862027168,
0.005127999931573868,
0.... | LLM Fine-Tuning, Quantization & Model Optimization | Large language models (LLMs) are increasingly used to provide conversational practice for English-as-a-second-language (ESL) learners. | We present TACT (Taxonomy-Aligned Conversational Tutor), a human-grounded framework for post-training and evaluating pedagogically adaptive ESL tutors. | On TACTBench, a strategy-balanced diagnostic benchmark comprising 78 authentic tutoring contexts, TACTutor improves over its backbone by 20.30% and outperforms all evaluated proprietary baselines under the same protocol, while maintaining backbone performance on established external educational benchmarks; in a blinded... | [
"Large Language Model (LLM)",
"Knowledge Distillation"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:37:57.740350 |
2608.03929v2 | Latent Reward Registers for Diffusion Preference Alignment | 2026-08-04T17:00:52Z | [
"cs.LG",
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yuanshen Guan | 5 | [
"Yuanshen Guan",
"Zipeng Feng",
"Chengru Song",
"Zhiwei Xiong",
"Peiqin Sun"
] | [] | http://arxiv.org/abs/2608.03929v2 | DETECTED_UNVERIFIED | [
"https://github.com/Guanys-dar/latent-reward-register"
] | [
0
] | [
false
] | Aligning diffusion models with human preferences usually relies on a sparse terminal reward evaluated on the final generated samples, presenting a severe temporal credit-assignment challenge across the multi-step denoising process. We propose Latent Reward Registers, a mechanism that estimates terminal preference direc... | [
-0.025766000151634216,
-0.10198000073432922,
0.0029360000044107437,
0.003200999926775694,
0.017249999567866325,
0.09875699877738953,
0.03703000023961067,
-0.016800999641418457,
0.04833900183439255,
0.014593999832868576,
0.01729400083422661,
-0.06502699851989746,
-0.018890000879764557,
0.03... | [
-0.09038999676704407,
-0.1158590018749237,
0.027242999523878098,
0.014367000199854374,
0.011152000166475773,
0.016839999705553055,
-0.06443700194358826,
-0.06465599685907364,
0.0785909965634346,
0.009595000185072422,
-0.05212699994444847,
-0.017038000747561455,
-0.0359870009124279,
0.00851... | LLM Fine-Tuning, Quantization & Model Optimization | Aligning diffusion models with human preferences usually relies on a sparse terminal reward evaluated on the final generated samples, presenting a severe temporal credit-assignment challenge across the multi-step denoising process. | We propose Latent Reward Registers, a mechanism that estimates terminal preference directly from intermediate noisy latents by prepending learnable, position-free register tokens to the input sequence of a frozen Diffusion Transformer (DiT). | Furthermore, RG-OPD outperforms online reinforcement learning baselines while reducing GPU hours by up to 33x, and RGS establishes a new state-of-the-art among training-free methods, strictly enhancing both alignment and perceptual metrics. | [
"Reinforcement Learning (RL)",
"Small Language Model (SLM)",
"Transformer Architecture"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Repo link (+1) | 2026-08-10T16:38:02.035673 |
2608.03920v1 | Equivariant Music Transformer | 2026-08-04T16:51:58Z | [
"cs.SD",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Zixun Guo | 2 | [
"Zixun Guo",
"Simon Dixon"
] | [] | http://arxiv.org/abs/2608.03920v1 | NOT_DETECTED | [] | [] | [] | Humans recognize a musical passage even when it is shifted in time or transposed in pitch, indicating a notion of equivariance in the representation space. Our analysis, however, shows that standard music transformers map such time-shifted or pitch-transposed inputs onto uncorrelated representations: these models becom... | [
-0.07997400313615799,
0.009290000423789024,
0.0017099999822676182,
-0.10451100021600723,
-0.10099200159311295,
0.061774998903274536,
0.021011000499129295,
-0.0004830000107176602,
-0.0005939999828115106,
-0.06481900066137314,
0.05649000033736229,
-0.041891999542713165,
0.00019299999985378236,... | [
-0.01730000041425228,
-0.08236400038003922,
0.08732999861240387,
-0.04492500051856041,
-0.018070999532938004,
0.09664399921894073,
-0.025932999327778816,
-0.04235300049185753,
0.06320299953222275,
-0.0735979974269867,
-0.03783300146460533,
-0.013279999606311321,
0.006837000139057636,
-0.00... | LLM Fine-Tuning, Quantization & Model Optimization | Humans recognize a musical passage even when it is shifted in time or transposed in pitch, indicating a notion of equivariance in the representation space. | Our analysis, however, shows that standard music transformers map such time-shifted or pitch-transposed inputs onto uncorrelated representations: these models become progressively less equivariant as they scale in size or train longer. | Through both objective and subjective evaluations, EMT demonstrates superior equivariance and generative capability compared to data augmentation, feature engineering, and state-of-the-art (SOTA) baselines. | [
"Small Language Model (SLM)",
"Transformer Architecture"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:38:02.054360 |
2608.03913v1 | Sparse Weight Decomposition for Efficient Circuit Extraction | 2026-08-04T16:40:48Z | [
"cs.LG",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Chuanhao Yan | 7 | [
"Chuanhao Yan",
"Xuhan Huang",
"Yawen Duan",
"Zhenfei Yin",
"Hang Zhao",
"Bryan Dai",
"Jie Fu"
] | [] | http://arxiv.org/abs/2608.03913v1 | NOT_DETECTED | [] | [] | [] | Dense pretrained transformers do not naturally expose interpretable units for circuit extraction. Existing approaches obtain such units by learning auxiliary sparse representations or training sparse models, incurring substantial additional computation while potentially introducing a fidelity gap between the representa... | [
-0.10507699847221375,
0.10345199704170227,
0.016742000356316566,
0.010011999867856503,
0.036924999207258224,
-0.06026000156998634,
0.025967000052332878,
-0.004222000017762184,
-0.04095099866390228,
0.0018350000027567148,
-0.010828999802470207,
0.005348000209778547,
0.027036000043153763,
-0... | [
-0.07947900146245956,
0.007085000164806843,
0.005233000032603741,
0.0499269999563694,
0.04161800071597099,
-0.00014200000441633165,
-0.08505800366401672,
-0.02812799997627735,
-0.03434700146317482,
-0.05147499963641167,
-0.026864999905228615,
-0.002901999978348613,
0.01293299999088049,
0.0... | LLM Fine-Tuning, Quantization & Model Optimization | Dense pretrained transformers do not naturally expose interpretable units for circuit extraction. | Existing approaches obtain such units by learning auxiliary sparse representations or training sparse models, incurring substantial additional computation while potentially introducing a fidelity gap between the representation being analyzed and the original pretrained model. | We further show that SWD remains effective for full-model replacement of all attention and MLP weight matrices after fine-tuning the nonzero factor values. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:38:02.073941 |
2608.03887v1 | Omega-S: A Functional Resilience Index for LLM Fine-Tuning | 2026-08-04T16:22:30Z | [
"cs.LG",
"cs.NE",
"q-bio.MN"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Alberto Acedo | 1 | [
"Alberto Acedo"
] | [] | http://arxiv.org/abs/2608.03887v1 | NOT_DETECTED | [] | [] | [] | Fine-tuning a large language model on new data degrades what it previously learned. We present Omega-S, a drop-in penalty computed from the weight matrix alone: it needs no previous-task data, no Fisher matrix and no stored copy of the old weights. It is three lines in an existing training loop and adds under 4% to the... | [
0.06493599712848663,
-0.06100200116634369,
0.01806900091469288,
0.023049000650644302,
-0.011896000243723392,
0.0066269999369978905,
-0.008821000345051289,
0.07605800032615662,
-0.07194600254297256,
-0.056373998522758484,
-0.07193099707365036,
0.042552001774311066,
0.03787299990653992,
-0.0... | [
-0.03186199814081192,
-0.00033999999868683517,
-0.03142099827528,
0.052553001791238785,
0.0024059999268501997,
0.01321099977940321,
-0.060679998248815536,
0.0026650000363588333,
0.029826000332832336,
-0.04538799822330475,
-0.03120799921452999,
0.026481999084353447,
-0.01971600018441677,
-0... | LLM Fine-Tuning, Quantization & Model Optimization | Fine-tuning a large language model on new data degrades what it previously learned. | We present Omega-S, a drop-in penalty computed from the weight matrix alone: it needs no previous-task data, no Fisher matrix and no stored copy of the old weights. | Code, per-seed results and the full record of negative results are available. | [
"Large Language Model (LLM)",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:38:02.095598 |
2608.03884v1 | BanglaWild: An In-the-Wild Bengali Scene Text Recognition Benchmark for OCR and Vision-Language Models | 2026-08-04T16:20:53Z | [
"cs.CV",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Sadab Shiper | 4 | [
"Sadab Shiper",
"Tawsif Tashwar Dipto",
"Mir Md Inzamam",
"Eshat Tanzeem"
] | [] | http://arxiv.org/abs/2608.03884v1 | NOT_DETECTED | [] | [] | [] | In-the-wild Bengali scene text recognition is largely unmeasured: existing resources target handwritten documents or constrained sign-board parsing, report only aggregate edit-distance metrics, and evaluate either conventional OCR or VLMs, never both on the same in-the-wild data. To address this gap, we introduce BANGL... | [
-0.05129000172019005,
0.014305000193417072,
-0.05925799906253815,
-0.002563999965786934,
-0.0008900000248104334,
0.07057099789381027,
-0.03777400031685829,
-0.015759000554680824,
0.035753000527620316,
-0.0844929963350296,
-0.04926599934697151,
-0.04281099885702133,
0.014530000276863575,
0.... | [
-0.014810999855399132,
-0.01904500089585781,
-0.024013999849557877,
0.0004079999926034361,
-0.0006360000115819275,
0.020018000155687332,
-0.09069699794054031,
0.010293000377714634,
0.019899999722838402,
-0.059268999844789505,
-0.04947299882769585,
-0.024351999163627625,
0.025693999603390694,... | LLM Fine-Tuning, Quantization & Model Optimization | In-the-wild Bengali scene text recognition is largely unmeasured: existing resources target handwritten documents or constrained sign-board parsing, report only aggregate edit-distance metrics, and evaluate either conventional OCR or VLMs, never both on the same in-the-wild data. | To address this gap, we introduce BANGLAWILD, a benchmark of 2,535 Bengali scene text images, each paired with a verbatim gold transcription, two categorical axes, four diagnostic attributes, and an orthographically standard form where the in-image text deviates from canonical spelling. | Prompt language mainly affects cross-script drift and LoRA reduces catastrophic failures in weak models without lifting the ceiling on already competent ones. | [
"Large Language Model (LLM)",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:38:02.116934 |
2608.04074v1 | Spend Bits Where Queries Look: KV Cache Vector Quantization with Attention-Preserving Transforms | 2026-08-04T16:10:59Z | [
"cs.LG",
"cs.AI",
"cs.IT",
"eess.SP"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Samuel Fernández-Menduiña | 5 | [
"Samuel Fernández-Menduiña",
"Amir Ziashahabi",
"Eduardo Pavez",
"Antonio Ortega",
"Salman Avestimehr"
] | [] | http://arxiv.org/abs/2608.04074v1 | NOT_DETECTED | [] | [] | [] | Long-context LLM decoding reads the key-value (KV) cache at every step. Loading it takes longer than computing attention over it, so throughput is bandwidth-bound. Hence, reducing the cache size can raise both decoding speed and serving capacity. The challenge is to reduce cache size while preserving the attention prod... | [
0.03538599982857704,
0.014371000230312347,
-0.0127379996702075,
0.03774299845099449,
-0.02674099989235401,
0.007780000101774931,
0.07803399860858917,
-0.058701999485492706,
0.07601699978113174,
0.0017399999778717756,
0.008307999931275845,
-0.03868500143289566,
-0.03043700009584427,
-0.0337... | [
0.0010809999657794833,
0.016186000779271126,
-0.027775999158620834,
-0.020176999270915985,
-0.028588000684976578,
-0.03528900071978569,
-0.0001880000054370612,
-0.02225000038743019,
0.06490500271320343,
0.021503999829292297,
-0.02269200049340725,
0.06246500089764595,
0.01926499977707863,
-... | LLM Fine-Tuning, Quantization & Model Optimization | Long-context LLM decoding reads the key-value (KV) cache at every step. | At two bits per element, the most competitive methods rely on orthogonal transforms. | At two bits per element, our method, termed NOVA-KV, recovers most of the long-context retrieval accuracy lost by scalar quantization methods at comparable throughput. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Quantization"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:38:02.133690 |
2608.03854v2 | Quantization Effects on Biomedical LLM Reliability | 2026-08-04T15:57:21Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Anton Rasmussen | 2 | [
"Anton Rasmussen",
"Hong Qin"
] | [] | http://arxiv.org/abs/2608.03854v2 | NOT_DETECTED | [] | [] | [] | When decoder language models are used as classifiers, predicted class probabilities depend on implementation choices, including the prompt template, verbalizer (label-to-token mapping), and scoring rule, that are rarely treated as experimental variables. We present a controlled evaluation of three Mistral-7B variants (... | [
0.01597600057721138,
-0.06667699664831161,
0.04297899827361107,
-0.03247400000691414,
-0.03383399918675423,
-0.0017079999670386314,
-0.016992000862956047,
0.04383999854326248,
0.02019999921321869,
-0.030551999807357788,
0.04508500173687935,
0.02822900004684925,
0.028687000274658203,
0.0401... | [
0.02889000065624714,
-0.029857000336050987,
0.012846999801695347,
-0.038711000233888626,
0.08425699919462204,
0.001370999962091446,
-0.040366001427173615,
0.08874599635601044,
0.06457500159740448,
-0.011130999773740768,
-0.0895640030503273,
-0.057787999510765076,
-0.00029299999005161226,
0... | LLM Fine-Tuning, Quantization & Model Optimization | When decoder language models are used as classifiers, predicted class probabilities depend on implementation choices, including the prompt template, verbalizer (label-to-token mapping), and scoring rule, that are rarely treated as experimental variables. | We present a controlled evaluation of three Mistral-7B variants (Base, BioMistral, and Instruct) on PubMed RCT sentence classification (n=2000) under FP16, INT8, and INT4 precision using four answer-text prompt templates. | These results demonstrate that prompt template design and scoring normalization are first-order experimental decisions when evaluating decoder language model calibration. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Quantization"
] | [
"82.7% accuracy"
] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:38:02.154995 |
2608.03769v1 | MDLMPE: Distribution Aware Positional Encoding for Masked Diffusion Language Models | 2026-08-04T14:54:14Z | [
"cs.CL",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Tong Ling | 8 | [
"Tong Ling",
"Hang Lei",
"Feng Xiao",
"Changhui Sun",
"Jiahang Xie",
"Hao Liu",
"Lu Liu",
"Yanlong Du"
] | [] | http://arxiv.org/abs/2608.03769v1 | NOT_DETECTED | [] | [] | [] | Masked diffusion language models (MDLMs) enable parallel generation and bidirectional context modeling, but their positional context differs fundamentally from that of autoregressive (AR) models. Whereas AR decoding exposes a contiguous prefix, MDLM denoising produces dynamic, non-contiguous configurations of revealed ... | [
-0.006004000082612038,
-0.12745900452136993,
0.00482300017029047,
-0.004951999988406897,
0.03180500119924545,
0.03850699961185455,
0.011518999934196472,
-0.09054899960756302,
0.10567399859428406,
-0.05209200084209442,
0.05353600159287453,
-0.022755999118089676,
0.04412499815225601,
0.03513... | [
-0.034373000264167786,
-0.13547299802303314,
0.024320999160408974,
0.0013970000436529517,
0.023110000416636467,
0.018837999552488327,
-0.06965900212526321,
-0.09954600036144257,
0.10276299715042114,
-0.09091100096702576,
0.02728400006890297,
0.010703000240027905,
0.04617699980735779,
0.015... | LLM Fine-Tuning, Quantization & Model Optimization | Masked diffusion language models (MDLMs) enable parallel generation and bidirectional context modeling, but their positional context differs fundamentally from that of autoregressive (AR) models. | To address this limitation, we propose MDLMPE, a positional encoding designed specifically for masked diffusion. | These results establish the evolving token-availability distribution as a useful positional signal for masked diffusion language models. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:38:02.200612 |
2608.03720v1 | Detecting Hallucinations and Recovering Verified Answers in Arabic Islamic Question Answering | 2026-08-04T14:19:53Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Khaled Ziani | 1 | [
"Khaled Ziani"
] | [] | http://arxiv.org/abs/2608.03720v1 | NOT_DETECTED | [] | [] | [] | Large language models can generate fluent responses to Islamic questions while introducing factual errors that are difficult to identify. This paper presents our system for \textsc{HalluScoring 2026} Task 2.1, \textit{Islamic Hallucination Detection and Find the Truth}. The task requires a unified two-step prediction: ... | [
-0.00030300000798888505,
0.0970269963145256,
-0.03346800059080124,
0.003982999827712774,
-0.03884800150990486,
0.04436499997973442,
0.06095699965953827,
-0.10975100100040436,
0.037108998745679855,
-0.052480001002550125,
-0.05651700124144554,
-0.05756400153040886,
0.07609699666500092,
0.018... | [
-0.013749999925494194,
0.07188399881124496,
-0.010923000052571297,
0.001218999968841672,
0.029232999309897423,
0.019142000004649162,
0.0039090001955628395,
-0.07329999655485153,
-0.024682000279426575,
-0.004941000137478113,
-0.07364899665117264,
-0.1260409951210022,
0.07455399632453918,
0.... | LLM Fine-Tuning, Quantization & Model Optimization | Large language models can generate fluent responses to Islamic questions while introducing factual errors that are difficult to identify. | This paper presents our system for \textsc{HalluScoring 2026} Task 2.1, \textit{Islamic Hallucination Detection and Find the Truth}. | These results yield a combined score of 0.912, demonstrating strong performance across both stages of the task. | [
"Large Language Model (LLM)"
] | [
"accuracy of 0.935",
"accuracy of 0.895"
] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:02.220617 |
2608.03632v1 | When Teachers Mislead: Spurious-Signal-Aware On-Policy Distillation | 2026-08-04T13:17:30Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yinuo Jiang | 7 | [
"Yinuo Jiang",
"Yongjie Ye",
"Zhou Tao",
"Xiang Zhuang",
"Qiang Zhang",
"Huajun Chen",
"Tiankai Li"
] | [] | http://arxiv.org/abs/2608.03632v1 | NOT_DETECTED | [] | [] | [] | On-Policy distillation (OPD) transfers teacher capabilities by supervising student-sampled trajectories with dense token-level teacher signals. Recent selective OPD methods improve this process by prioritizing signals that are confident, informative, or learnable. However, the assumptions overlook a fundamental failure... | [
-0.02013299986720085,
-0.0388059988617897,
0.11180000007152557,
0.0075610000640153885,
0.03340499848127365,
-0.07539299875497818,
0.05954600125551224,
-0.03715699911117554,
0.037687998265028,
0.011199000291526318,
-0.009891999885439873,
0.08853399753570557,
-0.02928999997675419,
0.01261300... | [
0.027605999261140823,
-0.12985099852085114,
0.15136699378490448,
0.030680999159812927,
0.06241700053215027,
-0.0516589991748333,
0.024176999926567078,
0.004964000079780817,
0.04990600049495697,
-0.03999900072813034,
-0.06924699991941452,
0.01828099973499775,
-0.00292299990542233,
0.0433689... | LLM Fine-Tuning, Quantization & Model Optimization | On-Policy distillation (OPD) transfers teacher capabilities by supervising student-sampled trajectories with dense token-level teacher signals. | Recent selective OPD methods improve this process by prioritizing signals that are confident, informative, or learnable. | These results establish input-groundedness as a key dimension for OPD supervision selection and offer a simple, effective strategy for mitigating spurious updates. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:02.237294 |
2608.03610v1 | Language-Specialized Multi-Teacher On-Policy Distillation for Multilingual LLM-Based ASR | 2026-08-04T13:02:47Z | [
"cs.CL",
"cs.SD",
"eess.AS"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yuan Xie | 6 | [
"Yuan Xie",
"Jiaqi Song",
"Xianliang Wang",
"Ming Lei",
"Jie Gao",
"Jie Wu"
] | [] | http://arxiv.org/abs/2608.03610v1 | NOT_DETECTED | [] | [] | [] | Modern LLM-based ASR systems have established multilingual capability as a standard feature, leveraging large-scale multilingual corpora and LLMs' cross-lingual knowledge to achieve competitive performance across multilingual benchmarks. However, joint modeling of languages with heterogeneous acoustic, phonological, an... | [
-0.06308799982070923,
-0.09837900102138519,
0.05716099962592125,
-0.01925400085747242,
-0.026301000267267227,
-0.03237299993634224,
0.0017369999550282955,
-0.04698999971151352,
-0.021310999989509583,
-0.05757400020956993,
-0.038385000079870224,
0.009344999678432941,
0.011505999602377415,
-... | [
0.0022279999684542418,
-0.13428500294685364,
0.05615299940109253,
-0.0003600000054575503,
-0.05468900129199028,
-0.011807999573647976,
-0.04641199856996536,
-0.035312000662088394,
-0.021170999854803085,
-0.06632799655199051,
-0.019284000620245934,
-0.05693599954247475,
0.005539000034332275,
... | LLM Fine-Tuning, Quantization & Model Optimization | Modern LLM-based ASR systems have established multilingual capability as a standard feature, leveraging large-scale multilingual corpora and LLMs' cross-lingual knowledge to achieve competitive performance across multilingual benchmarks. | However, joint modeling of languages with heterogeneous acoustic, phonological, and lexical characteristics inevitably introduces optimization conflicts, undermining language-wise specialization. | Experiments on benchmarks covering Mandarin, Mandarin subdialects, Cantonese, and English demonstrate that LS-MOPD substantially outperforms RL baselines and consistently surpasses the empirical performance envelope defined by best-performing RL teachers, revealing its potential to generalize beyond all teachers in mul... | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Blockchain/DeFi (+2) | 2026-08-10T16:38:02.260557 |
2608.03606v1 | Learning Clinical-Trial Strategy: Offline Policy Training for Decision Agents | 2026-08-04T12:58:17Z | [
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | William Bolton | 2 | [
"William Bolton",
"Philip Torr"
] | [] | http://arxiv.org/abs/2608.03606v1 | NOT_DETECTED | [] | [] | [] | Clinical development is sequential decision-making under uncertainty, where a sponsor must plan a portfolio of experiments from heterogeneous evidence. We study this setting by framing oncology clinical development as an offline decision-making problem in which an agent predicts the next six-month trial portfolio of an... | [
0.024311000481247902,
-0.006384999956935644,
-0.04159000143408775,
-0.0037299999967217445,
0.07851599901914597,
0.021640999242663383,
-0.03367900103330612,
0.10377000272274017,
0.012745999731123447,
0.03629700094461441,
-0.03339400142431259,
0.11140400171279907,
-0.060113999992609024,
0.00... | [
-0.0197759997099638,
0.006326999980956316,
-0.04377099871635437,
-0.021988999098539352,
0.041032999753952026,
0.0028029999230057,
-0.05818900093436241,
0.06676699966192245,
0.046107999980449677,
0.02327900007367134,
-0.05129000172019005,
0.041287001222372055,
-0.014144999906420708,
0.02705... | LLM Fine-Tuning, Quantization & Model Optimization | Clinical development is sequential decision-making under uncertainty, where a sponsor must plan a portfolio of experiments from heterogeneous evidence. | We study this setting by framing oncology clinical development as an offline decision-making problem in which an agent predicts the next six-month trial portfolio of an oncology drug program from information available at the decision date. | These results suggest that structured offline learning can teach agents to plan clinical experiments. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:02.283123 |
2608.03579v1 | Pin Once, Swap Light: Subspace-Aligned Centroid-Residual Training for Efficient Ultra-LoRA Serving | 2026-08-04T12:34:35Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Xiang Li | 4 | [
"Xiang Li",
"Pengcheng Wang",
"Huazheng Wang",
"Saurabh Bagchi"
] | [] | http://arxiv.org/abs/2608.03579v1 | NOT_DETECTED | [] | [] | [] | Modern multi-tenant Low-Rank Adapters (LoRAs) serving systems concurrently host tens to hundreds of LoRA adapters. Though powerful, this introduces a critical system dilemma between serving efficiency and task performance: higher-rank adapters generally achieve better downstream task performance, but their GPU VRAM foo... | [
-0.04716400057077408,
-0.1034879982471466,
0.008945000357925892,
0.01024600025266409,
-0.023700999096035957,
-0.027881000190973282,
-0.045974001288414,
-0.06228100135922432,
-0.015727000311017036,
-0.007216999772936106,
-0.03176100179553032,
0.07272599637508392,
-0.06158199906349182,
-0.04... | [
-0.013143000192940235,
-0.04879700019955635,
0.013119000010192394,
-0.02299400046467781,
0.038711000233888626,
-0.0047969999723136425,
-0.03409700095653534,
-0.03518899902701378,
-0.033330000936985016,
-0.03395000100135803,
-0.06934399902820587,
0.012783000245690346,
-0.0053400001488626,
-... | LLM Fine-Tuning, Quantization & Model Optimization | Modern multi-tenant Low-Rank Adapters (LoRAs) serving systems concurrently host tens to hundreds of LoRA adapters. | Though powerful, this introduces a critical system dilemma between serving efficiency and task performance: higher-rank adapters generally achieve better downstream task performance, but their GPU VRAM footprint and Host-to-Device PCIe swapping overhead severely constrain scalability. | When integrated into vLLM, SALT improves serving throughput by up to 51% under PCIe bandwidth pressure and 28% under GPU VRAM constraints for Llama-3.2-3B. | [
"Large Language Model (LLM)",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:38:02.304963 |
2608.03527v1 | Training Documents Reranker with Search Rubrics for Deep Research Agent | 2026-08-04T12:11:14Z | [
"cs.IR",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Wenhan Liu | 11 | [
"Wenhan Liu",
"Yu Lu",
"Qiaolin Xia",
"Hui Xu",
"Tong Zhao",
"Jian Xi",
"Yutao Zhu",
"Haijin Liang",
"Haibo Shi",
"Hao Wang",
"Zhicheng Dou"
] | [] | http://arxiv.org/abs/2608.03527v1 | NOT_DETECTED | [] | [] | [] | Retrieval systems help deep research agents generate high-quality answers by providing relevant documents. However, existing retrievers typically select documents through relevance matching, while individually well-matched top-$k$ documents may not form a \textit{set} that satisfies the complex information needs of an ... | [
-0.08757299929857254,
-0.06552000343799591,
-0.016227999702095985,
0.040268998593091965,
-0.010994999669492245,
0.06837599724531174,
-0.05331899970769882,
-0.013577999547123909,
-0.05793900042772293,
-0.042876001447439194,
-0.013007000088691711,
0.07655400037765503,
0.01601799950003624,
0.... | [
-0.08736299723386765,
-0.03031199984252453,
-0.02966099977493286,
0.03349899873137474,
-0.029803000390529633,
0.027744999155402184,
-0.000977999996393919,
0.017010999843478203,
-0.01246500015258789,
-0.030170999467372894,
-0.0527460016310215,
0.06237899884581566,
0.07792499661445618,
0.007... | LLM Fine-Tuning, Quantization & Model Optimization | Retrieval systems help deep research agents generate high-quality answers by providing relevant documents. | In this paper, we propose search-oriented rubrics that \textit{explicitly} define the requirements that high-quality document sets should satisfy for each agent query. | Extensive experiments demonstrate that RubricRanker outperforms the strongest baseline by 2.6 points on four deep research benchmarks and generalizes well to five RAG benchmarks. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)"
] | [
"outperforms the strongest baseline by 2.6"
] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Blockchain/DeFi (+2) | 2026-08-10T16:38:02.322996 |
2608.04056v1 | Learning Sexism Detection Using Multi-Agent Perspectivist Preference Optimization | 2026-08-04T11:35:30Z | [
"cs.CL",
"cs.CY",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Hadi Mohammadi | 5 | [
"Hadi Mohammadi",
"Tina Shahedi",
"Robert A. Bagheri",
"Mehdi Dastani",
"Masoume M. Raeissi"
] | [] | http://arxiv.org/abs/2608.04056v1 | NOT_DETECTED | [] | [] | [] | When people label text for sexism, they often disagree, and not because some of them are wrong: they genuinely perceive sexism differently. Most NLP systems discard this disagreement by collapsing it into a majority vote. We propose the Multi-Agent Perspectivist Preference Optimization (MAP-PO) framework to keep these ... | [
0.05685099959373474,
-0.06701699644327164,
-0.03237000107765198,
0.02248699963092804,
0.04195300117135048,
0.010579999536275864,
0.045354001224040985,
-0.08484800159931183,
-0.015158000402152538,
-0.04326700046658516,
0.04546000063419342,
-0.06760500371456146,
-0.012387000024318695,
0.0568... | [
0.07097700238227844,
-0.041652001440525055,
-0.02694600075483322,
0.00800899975001812,
0.04396500065922737,
0.065420001745224,
0.05703900009393692,
-0.03689200058579445,
-0.001307999948039651,
-0.04452500119805336,
0.042068000882864,
-0.08337400108575821,
0.011776000261306763,
0.0563430003... | LLM Fine-Tuning, Quantization & Model Optimization | When people label text for sexism, they often disagree, and not because some of them are wrong: they genuinely perceive sexism differently. | Most NLP systems discard this disagreement by collapsing it into a majority vote. | Second, we show that training each agent only on the labels of its own cluster pushes the agents far beyond the clusters they should represent, while adding a shared team-level training signal consistently keeps each agent calibrated to its cluster. | [
"Large Language Model (LLM)",
"Multi-Agent System"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Blockchain/DeFi (+2) | 2026-08-10T16:38:02.340949 |
2608.03457v1 | LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models | 2026-08-04T10:53:02Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Fengqi Zhu | 14 | [
"Fengqi Zhu",
"Shaoxuan Xu",
"Jingyang Ou",
"Zebin You",
"Yipeng Xing",
"Huabin Liu",
"Xiaolu Zhang",
"Jun Zhou",
"Zhenzhong Lan",
"Yankai Lin",
"Wayne Xin Zhao",
"Jianguo Li",
"Chongxuan Li",
"Ji-Rong Wen"
] | [] | http://arxiv.org/abs/2608.03457v1 | NOT_DETECTED | [] | [] | [] | Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE dLLMs, identifying qu... | [
0.017402000725269318,
-0.10602100193500519,
0.050415001809597015,
0.033649999648332596,
0.04592499881982803,
0.023607999086380005,
0.04686399921774864,
0.038913000375032425,
-0.001769999973475933,
-0.0907130017876625,
-0.0015180000336840749,
-0.06996600329875946,
-0.037491001188755035,
0.0... | [
-0.008985999971628189,
-0.12428099662065506,
-0.0042050001211464405,
0.06849300116300583,
-0.02326199971139431,
-0.023802999407052994,
-0.012710999697446823,
0.02583800069987774,
0.00949000008404255,
-0.013612999580800533,
-0.03374600037932396,
-0.03727100044488907,
-0.07085199654102325,
0... | LLM Fine-Tuning, Quantization & Model Optimization | Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. | We systematically characterize how optimization hyperparameters, compute allocation, and architecture scale for MoE dLLMs, identifying quantitative differences from scaling trends previously reported for AR models. | These results establish practical scaling laws and design principles for MoE dLLMs. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:02.366795 |
2608.03447v1 | Approximate Speculative Decoding | 2026-08-04T10:45:24Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yuannuo Feng | 8 | [
"Yuannuo Feng",
"Zegang Peng",
"Yuxin Xie",
"Yubing Ye",
"Yizhe Chen",
"Wenshuai Yao",
"Wenyong Zhou",
"Wang Kang"
] | [] | http://arxiv.org/abs/2608.03447v1 | VERIFIED_LIVE | [
"https://github.com/Kissmetothemoon/ASD"
] | [
200
] | [
true
] | Speculative decoding accelerates autoregressive generation by verifying a draft block with a target model in parallel. Under standard greedy verification, decoding stops at the first draft token that differs from the target argmax, discarding the remaining target-scored suffix. Although accepting such a mismatch change... | [
-0.052535999566316605,
-0.028212999925017357,
-0.009565000422298908,
-0.021602999418973923,
0.022874999791383743,
-0.037036001682281494,
0.04898399859666824,
-0.024630000814795494,
-0.0043299999088048935,
-0.024644000455737114,
-0.015205999836325645,
0.0154600003734231,
0.0029150000773370266... | [
-0.09364999830722809,
-0.04364300146698952,
0.03868899866938591,
0.010095999576151371,
0.05200999975204468,
-0.07101000100374222,
-0.03516500070691109,
-0.008333000354468822,
-0.007615999784320593,
-0.019388999789953232,
-0.05057799816131592,
0.0025710000190883875,
-0.041533999145030975,
-... | LLM Fine-Tuning, Quantization & Model Optimization | Speculative decoding accelerates autoregressive generation by verifying a draft block with a target model in parallel. | In this paper, we introduce \textbf{Approximate Speculative Decoding (ASD)}, a training-free verifier that replaces binary first-mismatch truncation with budgeted longest-prefix selection. | Experiments show that ASD improves fixed-workload throughput by $3.05\%$--$15.26\%$ over matched strict verification and averages a $7.78\%$ gain across seven Qwen3-14B + DSpark-14B tasks. | [
"Large Language Model (LLM)",
"Quantization",
"Speculative Decoding"
] | [] | Explosive (>50/mo) | 491 | 7 | Prototype (TRL 4-6) | Score 7/10. Live repo (+3) | 2026-08-10T16:38:03.520693 |
2608.03428v1 | OliveGemma: A 3 Billion Visual Language Model for Recognising the Mediterranean & European Diet | 2026-08-04T10:18:36Z | [
"cs.CV",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Dimitrios I. Zaridis | 9 | [
"Dimitrios I. Zaridis",
"Traianos Tsiokris",
"Vasileios C. Pezoulas",
"Daphni Plati",
"Eugenia Mylona",
"Eleni Georga",
"Nikos Tsiknakis",
"Antonis Sakellarios",
"Dimitrios I. Fotiadis"
] | [
"Hugging Face"
] | http://arxiv.org/abs/2608.03428v1 | VERIFIED_LIVE | [
"https://github.com/tsiokris/OliveGemma",
"https://huggingface.co/JamesZar/OliveGemma-3B"
] | [
200,
200
] | [
true,
true
] | Image based dietary assessment offers a scalable alternative to self reported food diaries, yet fine-grained food recognition remains challenging due to high intra-class variability and visually similar dishes. This study presents OliveGemma, a vision language model for recognising and reasoning about Mediterranean and... | [
0.05871199816465378,
-0.027104999870061874,
0.030837999656796455,
0.0021899999119341373,
0.0028059999458491802,
0.05443799868226051,
-0.02910899929702282,
-0.07696399837732315,
0.04987800121307373,
-0.05358099937438965,
0.04888800159096718,
-0.14139799773693085,
-0.053259000182151794,
-0.0... | [
0.0075159999541938305,
-0.07742299884557724,
0.02203500084578991,
-0.03474000096321106,
0.05982299894094467,
0.016395000740885735,
-0.021588999778032303,
-0.07066799700260162,
0.0036239998880773783,
-0.103022001683712,
-0.005863000173121691,
-0.09072399884462357,
-0.04213999956846237,
0.01... | LLM Fine-Tuning, Quantization & Model Optimization | Image based dietary assessment offers a scalable alternative to self reported food diaries, yet fine-grained food recognition remains challenging due to high intra-class variability and visually similar dishes. | This study presents OliveGemma, a vision language model for recognising and reasoning about Mediterranean and European cuisine. | The model is publicly available at https://huggingface.co/JamesZar/OliveGemma-3B and the experiments and results can be found at https://github.com/tsiokris/OliveGemma. | [
"Large Language Model (LLM)",
"LoRA / PEFT"
] | [
"accuracy of 92.96%"
] | Explosive (>50/mo) | 491 | 9 | Production-Ready (TRL 7-8) | Score 9/10. Live repo (+3); Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:38:04.481094 |
2608.06411v1 | Learning to Predict Middle-Layer Attention in MLLMs for Visual Token Prunin | 2026-08-04T05:44:07Z | [
"cs.AI",
"cs.CV"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Yuyao Sun | 6 | [
"Yuyao Sun",
"Tao Deng",
"Shuang Li",
"Deqing Wang",
"Hao Geng",
"Minjun Yu"
] | [] | http://arxiv.org/abs/2608.06411v1 | NOT_DETECTED | [] | [] | [] | Multimodal large language models (MLLMs) achieve strong performance across diverse vision-language tasks, but their efficiency is limited by the cost of processing numerous visual tokens. Visual token pruning can reduce this cost, but requires accurate token importance estimates. Recent studies have demonstrated that t... | [
0.03901499882340431,
-0.1048939973115921,
0.013508999720215797,
-0.020282000303268433,
0.09318500012159348,
0.04991700127720833,
0.027628999203443527,
0.0009449999779462814,
0.02476700022816658,
-0.010559000074863434,
0.04911400005221367,
-0.05048299953341484,
0.02447200007736683,
-0.00179... | [
0.06439000368118286,
-0.10110999643802643,
0.05689200013875961,
-0.008860000409185886,
0.15384599566459656,
0.05118599906563759,
-0.03273800015449524,
0.042371999472379684,
0.047717999666929245,
-0.025536999106407166,
-0.016982000321149826,
-0.062073998153209686,
-0.008280999958515167,
0.0... | LLM Fine-Tuning, Quantization & Model Optimization | Multimodal large language models (MLLMs) achieve strong performance across diverse vision-language tasks, but their efficiency is limited by the cost of processing numerous visual tokens. | Visual token pruning can reduce this cost, but requires accurate token importance estimates. | Recent studies have demonstrated that text-to-vision attention from middle language model layers can effectively guide visual token pruning, typically using attention from a predefined middle layer to select the visual tokens to retain. | [
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Blockchain/DeFi (+2) | 2026-08-10T16:38:04.503483 |
2608.03154v1 | ANCHOR-RE: An Agentic Neuro-Symbolic Framework for Grounded Biomedical Relation Extraction | 2026-08-04T05:36:19Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Shufan Ming | 5 | [
"Shufan Ming",
"Yikun Han",
"Gibong Hong",
"Rui Zhang",
"Halil Kilicoglu"
] | [] | http://arxiv.org/abs/2608.03154v1 | NOT_DETECTED | [] | [] | [] | Biomedical relation extraction (BioRE) extracts structured knowledge from biomedical literature for applications such as knowledge base construction and hypothesis generation. Traditional symbolic systems such as SemRep provide high precision but limited recall, while large language models (LLMs) offer stronger context... | [
0.0018939999863505363,
-0.0451119989156723,
0.057016000151634216,
-0.04798299819231033,
-0.03067000024020672,
0.031237000599503517,
-0.021364999935030937,
0.013233000412583351,
0.07644400000572205,
-0.07388599961996078,
-0.022199999541044235,
0.010064000263810158,
-0.015634000301361084,
0.... | [
-0.0268389992415905,
-0.04321400076150894,
0.014034999534487724,
0.01490699965506792,
0.056453999131917953,
-0.012107999995350838,
-0.10773900151252747,
0.06465400010347366,
0.009723000228404999,
-0.005268000066280365,
-0.07865700125694275,
0.026700999587774277,
0.04324299842119217,
0.0604... | LLM Fine-Tuning, Quantization & Model Optimization | Biomedical relation extraction (BioRE) extracts structured knowledge from biomedical literature for applications such as knowledge base construction and hypothesis generation. | Traditional symbolic systems such as SemRep provide high precision but limited recall, while large language models (LLMs) offer stronger contextual reasoning but remain prone to false-positive predictions. | Results across multiple benchmarks, model families, and post-cutoff literature support ANCHOR-RE as a practical training-free approach to biomedical literature mining. | [
"Large Language Model (LLM)",
"Multi-Agent System"
] | [
"precision of 69%"
] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:04.556768 |
2608.03137v1 | Verifiable Memory: Learning Unified Memory Management with Local and Global Verifiers for Large Language Model Agents | 2026-08-04T05:06:24Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Xiaolong Sun | 4 | [
"Xiaolong Sun",
"Qichao Wang",
"Hangyu Li",
"Liang Chen"
] | [] | http://arxiv.org/abs/2608.03137v1 | VERIFIED_LIVE | [
"https://github.com/Sun-SYSU-24/VerMem"
] | [
200
] | [
true
] | Large language model (LLM) agents must retain reusable information, control a bounded active context, and recover earlier evidence during long-horizon interaction. Existing methods commonly optimize long-term memory (LTM) and short-term memory (STM) separately, while unified policies are often trained primarily with tr... | [
-0.01832599937915802,
-0.04222400113940239,
-0.02820499986410141,
0.0036510000936686993,
0.044929999858140945,
0.022307999432086945,
-0.024879999458789825,
0.0243149995803833,
-0.004645999986678362,
0.006876000203192234,
0.01880900003015995,
-0.062178999185562134,
0.07265999913215637,
0.03... | [
0.03458800166845322,
-0.05168500170111656,
-0.0163199994713068,
0.018138999119400978,
0.09175000339746475,
0.09449499845504761,
-0.05515199899673462,
0.020642999559640884,
0.0066280001774430275,
-0.032210998237133026,
-0.014891000464558601,
0.0016060000052675605,
0.02705400064587593,
0.037... | LLM Fine-Tuning, Quantization & Model Optimization | Large language model (LLM) agents must retain reusable information, control a bounded active context, and recover earlier evidence during long-horizon interaction. | Existing methods commonly optimize long-term memory (LTM) and short-term memory (STM) separately, while unified policies are often trained primarily with trajectory-level feedback, which provides weak credit for individual memory decisions. | Under controlled online-token budgets on three interactive benchmarks, it also achieves the strongest efficiency--performance frontier among the compared methods. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 8 | Production-Ready (TRL 7-8) | Score 8/10. Live repo (+3); Benchmarks (+1) | 2026-08-10T16:38:05.373440 |
2608.03079v1 | CorePath: A Breast-Specialized Pathology Foundation Model for Core Needle Biopsy Diagnosis and Risk-Controlled Report Generation | 2026-08-04T03:43:49Z | [
"cs.CV",
"cs.AI",
"cs.LG",
"stat.AP"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Ting Yin | 20 | [
"Ting Yin",
"Danning Li",
"Chen Shu",
"Xiaoxia Yao",
"Boyu Fu",
"Yujing Chang",
"Tianyu Shi",
"Mengna Feng",
"Jie Chen",
"Jing Fu",
"Xiuli Xiao",
"Tianlin Li",
"Mumin Shao",
"Jiaxin Bi",
"Wenchuan Zhang",
"Xiaoyan Wu",
"Xiao Han",
"Zhang Zhang",
"Yuhao Yi",
"Hong Bu"
] | [] | http://arxiv.org/abs/2608.03079v1 | NOT_DETECTED | [] | [] | [] | Breast core needle biopsy (CNB) is central to breast cancer diagnosis yet remains challenging because limited tissue sampling, lesion heterogeneity, and subtle morphologic overlap can obscure subtype distinctions. We developed CorePath, a breast-specialized multimodal pathology foundation model fine-tuned from PRISM us... | [
0.035016998648643494,
-0.08117000013589859,
-0.04363499954342842,
0.0172520000487566,
0.04786600172519684,
-0.05650300160050392,
-0.0469839982688427,
0.07781700044870377,
-0.06577300280332565,
-0.019113000482320786,
-0.03900400176644325,
-0.05155299976468086,
0.023429999127984047,
0.095338... | [
0.07020799815654755,
-0.11833299696445465,
-0.02865000069141388,
-0.019185999408364296,
0.07795800268650055,
-0.07877500355243683,
-0.08315200358629227,
0.026883000507950783,
-0.08461999893188477,
-0.05023900046944618,
-0.08747900277376175,
-0.09436500072479248,
0.05006900057196617,
0.0714... | LLM Fine-Tuning, Quantization & Model Optimization | Breast core needle biopsy (CNB) is central to breast cancer diagnosis yet remains challenging because limited tissue sampling, lesion heterogeneity, and subtle morphologic overlap can obscure subtype distinctions. | We developed CorePath, a breast-specialized multimodal pathology foundation model fine-tuned from PRISM using 7901 paired CNB whole-slide images and diagnostic reports from two centers. | These results demonstrate that domain-specialized foundation models with statistical risk control offer a promising approach for accurate breast CNB diagnosis and reliable report generation. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:05.426688 |
2608.03071v1 | Getting the Parameters Right: A Difficulty-Graded Benchmark and Probe-Guided Training for LLM Tool Calls | 2026-08-04T03:36:41Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Guoyao Yu | 16 | [
"Guoyao Yu",
"Xiaoqing Sun",
"Ziqi Huang",
"Shaojing Fan",
"Zhongyi Zhang",
"Xiaomeng Hu",
"Xiaobo Xue",
"Yangyang Shi",
"Xiong Xiao",
"Yang Song",
"Biao Lyu",
"Rong Wen",
"Xing Li",
"Qinming He",
"Shunming Zhu",
"Zhenguang Liu"
] | [] | http://arxiv.org/abs/2608.03071v1 | NOT_DETECTED | [] | [] | [] | Large language model agents derive much of their capability from tool use. Existing research on tool use has largely focused on selecting the right tool and orchestrating the order of calls. However, correctly filling the parameters of a tool call is equally critical for successful execution and has received far less a... | [
-0.040481001138687134,
-0.027132000774145126,
0.016481000930070877,
0.029451999813318253,
-0.05182600021362305,
-0.07940100133419037,
-0.015608999878168106,
0.017271999269723892,
-0.07755299657583237,
-0.05136900022625923,
-0.07827000319957733,
-0.08984299749135971,
-0.004362000152468681,
... | [
-0.04439299926161766,
-0.1079770028591156,
0.015367000363767147,
0.053004998713731766,
-0.02575100027024746,
-0.01586199924349785,
-0.05616400018334389,
-0.00687299994751811,
0.05107000097632408,
-0.01081900019198656,
-0.07228799909353256,
-0.039014000445604324,
0.038867998868227005,
0.004... | LLM Fine-Tuning, Quantization & Model Optimization | Large language model agents derive much of their capability from tool use. | Existing research on tool use has largely focused on selecting the right tool and orchestrating the order of calls. | Extensive experiments across 5 open models on ParamBench and 6 external benchmarks demonstrate that our method substantially improves parameter generation, raising the average exact match from 19.7% to 59.6%. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:05.493972 |
2608.03063v1 | SeqLLM: Augmenting LLMs with Behavioral-Sequence Modeling for High-Stakes Decisions at WeChat Pay | 2026-08-04T03:23:12Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Guilin Li | 5 | [
"Guilin Li",
"Jiaxing Zhang",
"Matthias Hwai Yong Tan",
"Bo Wang",
"Weiran Huang"
] | [] | http://arxiv.org/abs/2608.03063v1 | NOT_DETECTED | [] | [] | [] | Merchant risk control at large payment platforms screens tens of millions of merchants daily, where false positives harm legitimate merchants and false negatives leave harmful activity undetected. The hardest cases require jointly understanding a merchant's textual profile and long behavioral sequence. Large language m... | [
-0.03608600050210953,
0.007007999811321497,
0.013344000093638897,
-0.010197999887168407,
-0.024416999891400337,
0.010714000090956688,
0.06756199896335602,
-0.032134998589754105,
0.037654001265764236,
0.008887999691069126,
-0.03586599975824356,
-0.016839999705553055,
-0.021309999749064445,
... | [
-0.027535000815987587,
-0.03486799821257591,
-0.033920999616384506,
0.01785000041127205,
-0.029774999246001244,
0.06357099860906601,
0.03205300122499466,
-0.005423000082373619,
0.026518000289797783,
-0.09835200011730194,
0.0038139999378472567,
-0.019889000803232193,
0.07872600108385086,
-0... | LLM Fine-Tuning, Quantization & Model Optimization | Merchant risk control at large payment platforms screens tens of millions of merchants daily, where false positives harm legitimate merchants and false negatives leave harmful activity undetected. | We present SeqLLM, a framework that adds behavioral-sequence modeling to a pretrained LLM while preserving its language ability. | On RecIF, it improves Pass@32 by 14.2% over the full OneRec-8B pipeline using only one-fifth of its GPU-days. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:05.539676 |
2608.03036v1 | LLM Serving in the Wild: An Empirical Study of Frameworks, Methods, and System Designs | 2026-08-04T02:33:16Z | [
"cs.SE",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Forough Majidi | 4 | [
"Forough Majidi",
"Mohammad Mehdi Morovati",
"Foutse Khomh",
"Heng Li"
] | [] | http://arxiv.org/abs/2608.03036v1 | NOT_DETECTED | [] | [] | [] | Large Language Models (LLMs) are integrated into software systems and AI services, making efficient LLM serving a concern for software engineering. Serving LLMs is challenging because inference requires computation, memory, GPU resources, and execution while maintaining latency and throughput. Although prior research h... | [
0.07706999778747559,
-0.03114599920809269,
-0.009266999550163746,
-0.08004599809646606,
0.06740500032901764,
-0.037728000432252884,
-0.04374200105667114,
0.059532999992370605,
-0.030615000054240227,
0.04108700156211853,
-0.08004699647426605,
0.05344099923968315,
0.07336500287055969,
-0.025... | [
0.005305999889969826,
-0.10695499926805496,
-0.01386099960654974,
-0.0041589997708797455,
0.1024560034275055,
-0.011478000320494175,
-0.03921699896454811,
0.029899999499320984,
0.025373000651597977,
-0.03737799823284149,
-0.102572001516819,
0.022105000913143158,
0.0187120009213686,
0.00708... | LLM Fine-Tuning, Quantization & Model Optimization | Large Language Models (LLMs) are integrated into software systems and AI services, making efficient LLM serving a concern for software engineering. | Although prior research has proposed LLM inference, optimization, and serving techniques and frameworks, little is known about how they are adopted in practice. | Our results show that vLLM is the most visible framework in popularity and adoption, while parallel computation, memory management, and network pruning are the most frequently used serving-method categories. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:05.588629 |
2608.03020v2 | LoCA: Forward-Only LLM Tuning after One-Shot Calibration with Local Credit Assignment | 2026-08-04T02:06:43Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Linhan Xia | 6 | [
"Linhan Xia",
"Rui Liu",
"Zhaofeng Zhang",
"Yihao Wang",
"Binrui Shen",
"Shengxin Zhu"
] | [] | http://arxiv.org/abs/2608.03020v2 | VERIFIED_LIVE | [
"https://github.com/Xia12121/LoCA"
] | [
200
] | [
true
] | Parameter-efficient post-training reduces the number of trainable parameters, but still requires repeated end-to-end backpropagation through the frozen backbone. Every adaptation step therefore needs backward-capable hardware and must store or recompute activations. We ask whether this repeated backward chain can be re... | [
0.018091000616550446,
-0.044401999562978745,
-0.029503999277949333,
-0.028178999200463295,
0.054986998438835144,
-0.013325000181794167,
0.029207000508904457,
-0.06095600128173828,
0.002334000077098608,
0.0011439999798312783,
-0.020438000559806824,
-0.0410429984331131,
0.02129100076854229,
... | [
-0.07097800076007843,
-0.07123500108718872,
0.01978600025177002,
0.05559699982404709,
0.0899370014667511,
0.03060300089418888,
-0.05658600106835365,
-0.04551199823617935,
-0.03188600018620491,
-0.03256100043654442,
-0.01742500066757202,
-0.04153300076723099,
0.003208999987691641,
-0.009583... | LLM Fine-Tuning, Quantization & Model Optimization | Parameter-efficient post-training reduces the number of trainable parameters, but still requires repeated end-to-end backpropagation through the frozen backbone. | We introduce Local Credit Assignment (LoCA), a two-stage method for small-shift adaptation. | We evaluate LoCA on five discriminative benchmarks with Qwen2.5 models from 0.5B to 14B. | [
"Large Language Model (LLM)",
"Transformer Architecture",
"LoRA / PEFT"
] | [] | Explosive (>50/mo) | 491 | 9 | Production-Ready (TRL 7-8) | Score 9/10. Live repo (+3); Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:38:06.672965 |
2608.02975v1 | TQLite: Multi-LLM Jury Guided Distillation for Real-time MQM Translation Quality Evaluation | 2026-08-04T00:24:06Z | [
"cs.CL",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Bhavin Jawade | 2 | [
"Bhavin Jawade",
"Cameron R. Wolfe"
] | [] | http://arxiv.org/abs/2608.02975v1 | NOT_DETECTED | [] | [] | [] | Large language models (LLMs) have demonstrated impressive performance in MQM-based translation quality (TQ) evaluation, and recent advances in large reasoning models (LRMs) promise even greater improvements. However, both LLMs and LRMs are computationally expensive to deploy at scale, while small language models (SLMs)... | [
0.0017729999963194132,
-0.05786899849772453,
0.061730001121759415,
-0.03296099975705147,
-0.08670300245285034,
-0.05293799936771393,
-0.02592100016772747,
-0.043157000094652176,
0.03432900086045265,
-0.04597200080752373,
-0.06830500066280365,
0.00827299989759922,
0.02528100088238716,
-0.03... | [
0.02104399912059307,
-0.03268999978899956,
0.06270000338554382,
0.02049200050532818,
0.05117899924516678,
-0.059967998415231705,
-0.024403000250458717,
0.005801999941468239,
0.014032999984920025,
-0.014077000319957733,
-0.05572599917650223,
-0.059147000312805176,
0.03298800066113472,
0.047... | LLM Fine-Tuning, Quantization & Model Optimization | Large language models (LLMs) have demonstrated impressive performance in MQM-based translation quality (TQ) evaluation, and recent advances in large reasoning models (LRMs) promise even greater improvements. | In this work, we present an extensive empirical study benchmarking SLMs, LLMs, and LRMs across a wide range of TQ evaluation setups, providing a comprehensive view of the current landscape and establishing best practices. | Our results demonstrate that SLMs trained via TQLite achieve strong MQM evaluation performance that far exceeds off-the-shelf evaluation capabilities of standard SLMs, offering a scalable and cost-effective alternative to LLM- and LRM-based evaluators. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:06.701789 |
2608.02961v1 | Scaling an Autoregressive Transformer for Single-Cell Generation | 2026-08-03T23:54:30Z | [
"cs.LG",
"cs.AI",
"q-bio.GN"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Aleksandr Sharipov | 3 | [
"Aleksandr Sharipov",
"Yusif Mukhtarov",
"Igor Molybog"
] | [] | http://arxiv.org/abs/2608.02961v1 | NOT_DETECTED | [] | [] | [] | We study a self-supervised generation task for single-cell gene expression vectors: given a set of vectors from a cell type, we aim to generate additional gene expression vectors of that cell type. For this task we characterize both the biological fidelity of the generated gene expression vectors and the scaling behavi... | [
-0.15433800220489502,
0.0342629998922348,
-0.016791999340057373,
0.01839200034737587,
-0.09872200340032578,
0.0048779998905956745,
-0.06459099799394608,
0.02202400006353855,
0.007805999834090471,
-0.03556999936699867,
-0.006010999903082848,
-0.0731159970164299,
-0.027612000703811646,
-0.05... | [
-0.07671499997377396,
-0.07527799904346466,
0.05861499905586243,
0.030563000589609146,
0.013287000358104706,
0.03491000086069107,
-0.009360999800264835,
0.015066999942064285,
0.08236700296401978,
-0.015661999583244324,
-0.05901600047945976,
-0.08769699931144714,
-0.021841000765562057,
-0.0... | LLM Fine-Tuning, Quantization & Model Optimization | We study a self-supervised generation task for single-cell gene expression vectors: given a set of vectors from a cell type, we aim to generate additional gene expression vectors of that cell type. | We study the scaling properties of the proposed architecture by varying the number of trained parameters and the amount of training data. | To our knowledge, we find the first jointly-fit two-exponent scaling law and compute-optimal frontier for a single-cell foundation model. | [
"Small Language Model (SLM)",
"Transformer Architecture",
"Quantization"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:38:06.725597 |
2608.02951v1 | SP3O: Reinforcement Learning from Segment Preferences without Reward Modeling | 2026-08-03T23:28:06Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Evan Assmus | 3 | [
"Evan Assmus",
"Qining Zhang",
"Lei Ying"
] | [] | http://arxiv.org/abs/2608.02951v1 | NOT_DETECTED | [] | [] | [] | Preference-based reinforcement learning (PbRL) for general stochastic MDPs often requires training a reward model. Existing reward-model-free methods are either restricted to bandits or deterministic MDPs, such as DPO or P3O, or use zeroth-order, gradient-free optimization, which in general exhibits a slower convergenc... | [
-0.06066500023007393,
-0.07516700029373169,
0.006862000096589327,
-0.041891999542713165,
-0.00797600019723177,
0.014666000381112099,
0.03634899854660034,
-0.005400000140070915,
-0.028049999848008156,
0.009317999705672264,
-0.016262000426650047,
-0.059898000210523605,
0.029051000252366066,
... | [
-0.03490599989891052,
-0.07179900258779526,
0.0009399999980814755,
-0.05106600001454353,
0.0194109994918108,
0.04039600118994713,
-0.039361000061035156,
0.042454998940229416,
0.07024399936199188,
0.02645999938249588,
-0.04787300154566765,
0.0716560035943985,
0.033594999462366104,
0.0376730... | LLM Fine-Tuning, Quantization & Model Optimization | Preference-based reinforcement learning (PbRL) for general stochastic MDPs often requires training a reward model. | Existing reward-model-free methods are either restricted to bandits or deterministic MDPs, such as DPO or P3O, or use zeroth-order, gradient-free optimization, which in general exhibits a slower convergence rate than gradient-based algorithms. | We also evaluate it experimentally against other PbRL/RLHF algorithms in robotic control and LLM finetuning settings to show its improved performance, especially in long-horizon tasks. | [
"Reinforcement Learning (RL)",
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:38:06.751168 |
2608.02948v1 | Rubrics as Privileged Information for Open-Ended Generation | 2026-08-03T23:25:50Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Deepika Bablani | 3 | [
"Deepika Bablani",
"Ajay Gupta",
"Wanming Chen"
] | [] | http://arxiv.org/abs/2608.02948v1 | NOT_DETECTED | [] | [] | [] | On-policy self-distillation (OPSD), where a single model acts as both student and teacher with different contexts, has shown promise in verifiable domains like math, where hard privileged information (PI) in the form of ground-truth answers structurally constrains valid continuations. We extend OPSD to open-ended gener... | [
-0.07675399631261826,
-0.03619600087404251,
-0.08787299692630768,
-0.004505999851971865,
0.016183000057935715,
0.006732999812811613,
-0.01449500024318695,
-0.005766000133007765,
-0.05111699923872948,
-0.028807999566197395,
0.016426000744104385,
0.014643000438809395,
0.048496998846530914,
-... | [
-0.11266700178384781,
-0.06339599937200546,
0.04277399927377701,
0.0011660000309348106,
-0.011776000261306763,
0.002589999930933118,
0.0023950000759214163,
0.04259800165891647,
0.02155900001525879,
-0.009584999643266201,
-0.05294499918818474,
0.05148499831557274,
0.007085000164806843,
0.01... | LLM Fine-Tuning, Quantization & Model Optimization | On-policy self-distillation (OPSD), where a single model acts as both student and teacher with different contexts, has shown promise in verifiable domains like math, where hard privileged information (PI) in the form of ground-truth answers structurally constrains valid continuations. | We extend OPSD to open-ended generation using soft PI in the form of rubrics that guide preferences but admit many valid responses. | We further show that these findings generalize to training on the RubricHub Science corpus and evaluating on ResearchQA: soft rubric PI outperforms both reference-PI distillation and RaR RL (66.6% vs. 64.2% and 57.6%). | [
"Reinforcement Learning (RL)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:06.885153 |
2608.02947v1 | ATFlash: Per-RoPE-Wavelength Attention Windows for Compute/Memory-Efficient LLM Inference | 2026-08-03T23:23:38Z | [
"cs.LG",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Shun-ichiro Hayashi | 4 | [
"Shun-ichiro Hayashi",
"Daichi Mukunoki",
"Tetsuya Hoshino",
"Takahiro Katagiri"
] | [
"OpenAI"
] | http://arxiv.org/abs/2608.02947v1 | NOT_DETECTED | [] | [] | [] | The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position. Aligned with this structure, we propose the per-RoPE-wavelength distance window: it prunes the query--key inner-product t... | [
0.02216700091958046,
-0.09816300123929977,
0.0004799999878741801,
0.008750000037252903,
0.01109199970960617,
-0.02223300002515316,
0.023910000920295715,
-0.025319000706076622,
-0.022106999531388283,
-0.006378000136464834,
0.01625099964439869,
-0.08696699887514114,
-0.039872001856565475,
0.... | [
-0.01606000028550625,
-0.0335640013217926,
-0.040130000561475754,
-0.06379099935293198,
-0.01357400044798851,
0.030208999291062355,
0.018962999805808067,
0.008957000449299812,
0.056752998381853104,
-0.01936900056898594,
0.012632000260055065,
0.039267998188734055,
0.020695999264717102,
-0.0... | LLM Fine-Tuning, Quantization & Model Optimization | The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position. | Aligned with this structure, we propose the per-RoPE-wavelength distance window: it prunes the query--key inner-product terms beyond a wavelength-proportional distance. | End to end on Qwen2.5-7B-1M, with 57\% of the inner-product terms pruned, the speedup reaches $1.31\times$ at a 1M-token context. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Flash Attention"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Benchmarks (+1) | 2026-08-10T16:38:06.921270 |
2608.02940v2 | When Compression Scores Cannot Decide: Information Boundaries for Group-Robust LLM Pruning | 2026-08-03T23:07:29Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Andrew Zhang | 1 | [
"Andrew Zhang"
] | [] | http://arxiv.org/abs/2608.02940v2 | NOT_DETECTED | [] | [] | [] | A stable compression score can still select the worse model. In our dense study, a split-half reliable path-quadratic score predicted a 16.1\% gain, while the selected endpoints were 6.0--7.7% worse than two controls. We ask what a compression statistic can justify when deployment cares about the worst supplied group. ... | [
-0.01993199996650219,
0.07350599765777588,
-0.060837000608444214,
-0.010351999662816525,
0.06853300333023071,
0.015402000397443771,
-0.009859000332653522,
-0.04526200145483017,
-0.023194000124931335,
-0.03290000185370445,
0.00023799999326001853,
0.06373000144958496,
0.09621000289916992,
-0... | [
0.005834999959915876,
-0.08240900188684464,
0.06733900308609009,
0.09668900072574615,
0.042374998331069946,
-0.03309300169348717,
-0.07637900114059448,
0.01933100074529648,
0.04317900165915489,
0.01728600077331066,
-0.03611399978399277,
0.05282299965620041,
0.02931700088083744,
-0.02476000... | LLM Fine-Tuning, Quantization & Model Optimization | A stable compression score can still select the worse model. | In our dense study, a split-half reliable path-quadratic score predicted a 16.1\% gain, while the selected endpoints were 6.0--7.7% worse than two controls. | A compute-matched hard-max trajectory ends 32.7% worse than pooled, and neither adaptive trajectory improves excess NLL. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)"
] | [] | Explosive (>50/mo) | 491 | 4 | Early Research (TRL 1-3) | Score 4/10. | 2026-08-10T16:38:06.933075 |
2608.02901v1 | AnchorKV: Anchor-Residual KV Cache Compression | 2026-08-03T21:38:30Z | [
"cs.LG",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Malik Khalaf | 5 | [
"Malik Khalaf",
"Yara Shamshoum",
"Nitzan Hodos",
"Yuval Sieradzki",
"Assaf Schuster"
] | [] | http://arxiv.org/abs/2608.02901v1 | NOT_DETECTED | [] | [] | [] | The key-value (KV) cache is the primary memory bottleneck in long-context LLM inference. Existing approaches attack it from opposite ends: eviction methods permanently discard tokens, degrading performance whenever a discarded token later proves essential, while quantization methods retain all tokens at low precision b... | [
-0.0679280012845993,
0.11294800043106079,
-0.01972999982535839,
0.029395999386906624,
0.03319400176405907,
-0.09121900051832199,
-0.035388000309467316,
-0.03980199992656708,
0.0012069999938830733,
0.023329999297857285,
-0.012466000393033028,
0.020862000063061714,
-0.0371830016374588,
-0.04... | [
0.001728000002913177,
0.03557100147008896,
-0.005028999876230955,
0.01953200064599514,
0.04619399830698967,
-0.014618000015616417,
-0.0230919998139143,
0.014310999773442745,
0.05376499891281128,
0.006012999918311834,
-0.016339000314474106,
0.03181200101971626,
0.008706999942660332,
-0.0389... | LLM Fine-Tuning, Quantization & Model Optimization | The key-value (KV) cache is the primary memory bottleneck in long-context LLM inference. | Existing approaches attack it from opposite ends: eviction methods permanently discard tokens, degrading performance whenever a discarded token later proves essential, while quantization methods retain all tokens at low precision but offer limited compression. | AnchorKV consistently preserves accuracy across models and datasets, retaining 99% of the full-cache score at the 70B scale, while keeping the entire context at a fraction of its cost. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Quantization"
] | [] | Explosive (>50/mo) | 491 | 6 | Prototype (TRL 4-6) | Score 6/10. Benchmarks (+1); Quantization tool (+1) | 2026-08-10T16:38:06.956329 |
2608.02843v1 | MutMem: Cryptographically Authorized Mutation in Persistent Agent Memory | 2026-08-03T19:58:37Z | [
"cs.CR",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Walid Saidi | 1 | [
"Walid Saidi"
] | [] | http://arxiv.org/abs/2608.02843v1 | NOT_DETECTED | [] | [] | [] | Persistent agent memory must adapt as later outcomes change earlier evidence, yet mutable retrieval weights create an attribution problem: reviewers must distinguish authorized adaptation from database tampering. We present MutMem, an authorized-mutation protocol in HOM-AIMOS, a persistent agent-memory engine. MutMem r... | [
-0.05428599938750267,
0.039051998406648636,
-0.09921599924564362,
0.00662099989131093,
-0.01802399940788746,
-0.04639799892902374,
0.015176000073552132,
0.008004000410437584,
0.043675001710653305,
-0.03576799854636192,
0.030664000660181046,
0.00798800028860569,
0.07425999641418457,
0.00590... | [
-0.062403999269008636,
0.023322999477386475,
-0.05793600156903267,
-0.00595899997279048,
0.07400599867105484,
-0.054225001484155655,
0.0330130010843277,
0.007797999773174524,
0.00019099999917671084,
-0.015935000032186508,
0.023412000387907028,
-0.008220000192523003,
0.11601399630308151,
-0... | LLM Fine-Tuning, Quantization & Model Optimization | Persistent agent memory must adapt as later outcomes change earlier evidence, yet mutable retrieval weights create an attribution problem: reviewers must distinguish authorized adaptation from database tampering. | We present MutMem, an authorized-mutation protocol in HOM-AIMOS, a persistent agent-memory engine. | We evaluate utility, mutation integrity, and poisoning adaptation. | [
"Large Language Model (LLM)",
"Small Language Model (SLM)",
"Quantization"
] | [] | Explosive (>50/mo) | 491 | 5 | Prototype (TRL 4-6) | Score 5/10. Quantization tool (+1) | 2026-08-10T16:38:06.985289 |
2608.02786v1 | Evaluation Blindness: How Silent Measurement Failures Corrupt AI Systems from Training to Deployment | 2026-08-03T18:33:37Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | Priyanka Bajaj | 1 | [
"Priyanka Bajaj"
] | [
"Independent Researcher"
] | http://arxiv.org/abs/2608.02786v1 | VERIFIED_LIVE | [
"https://github.com/priyanka25aug/llm-failure-taxonomy"
] | [
200
] | [
true
] | AI systems can fail silently. The failure propagates through training loops, evaluation pipelines, and production monitoring stacks until downstream harm makes it visible. This paper introduces evaluation blindness: a measurement function M exhibits evaluation blindness with respect to failure class F when it produces ... | [
0.001307999948039651,
-0.02670300006866455,
-0.004724999889731407,
0.06698200106620789,
0.03432900086045265,
-0.009819000028073788,
0.03745799884200096,
-0.042493000626564026,
-0.006076999939978123,
-0.06028499826788902,
-0.03993599861860275,
-0.048583000898361206,
0.049525998532772064,
-0... | [
-0.03293899819254875,
-0.025067999958992004,
-0.024135999381542206,
0.032885998487472534,
0.08040200173854828,
-0.015776999294757843,
0.024473000317811966,
-0.01218700036406517,
-0.03666999936103821,
0.01242899987846613,
-0.06404999643564224,
-0.07314900308847427,
0.03389599919319153,
-0.0... | LLM Fine-Tuning, Quantization & Model Optimization | AI systems can fail silently. | This paper introduces evaluation blindness: a measurement function M exhibits evaluation blindness with respect to failure class F when it produces readings indistinguishable from a healthy state while the system is actually failing, with no auxiliary signal flagging the gap. | Data, code, and taxonomy schema are at https://github.com/priyanka25aug/llm-failure-taxonomy. | [
"Large Language Model (LLM)"
] | [] | Explosive (>50/mo) | 491 | 9 | Production-Ready (TRL 7-8) | Score 9/10. Live repo (+3); Blockchain/DeFi (+2) | 2026-08-10T16:38:07.926073 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.