LoRA

2026-10-28

RESEARCH · 1 SOURCE · arXiv cs.AI

LoRA on a Budget — Functional Rank Allocation (FRA) for adapter-fleet compression

This arXiv preprint introduces Functional Rank Allocation (FRA), a post-hoc method that builds per-adapter functional rank–distortion frontiers and solves a fleet-level minimax allocation to compress LoRA adapters under a shared rank budget. FRA combines isolated-module measurements, a dynamic-programming search surrogate, end-to-end verification and a complementary spectral proposal; evaluated on 78 adapters from three public fleets it improves worst-case and tail utility versus uniform spectral compression in constrained-budget regimes.

6.0

2026-10-27

COMPANIES · 1 SOURCE · AWS Machine Learning

Exowatt uses SageMaker serverless customization to fine-tune Qwen3 models for edge and cloud, beating frontier models

Exowatt built a repeatable pipeline on Amazon SageMaker AI serverless model customization to produce task-specific fine-tuned models for edge alarm triage (Qwen3-1.7B) and cloud energy-market analysis (Qwen3-14B). The tuned models outperformed benchmark frontier models on accuracy for these tasks and reached up to 49× lower cost, while enabling a repeatable path from raw operational data to deployable artifacts in days rather than months.

6.0

2026-10-23

RESEARCH · 1 SOURCE · arXiv cs.AI

ClariEdge dataset and lightweight LoRA improve query clarification for RAG

The paper (arXiv:2610.28686v1) introduces ClariEdge, a ~3,900-example dataset of ambiguous and non-ambiguous conversational queries for retrieval-augmented generation (RAG). A lightweight LoRA fine-tune achieves 96.6% classification accuracy and outperforms prompted larger models on both ambiguity detection and quality of clarification requests; the dataset is published on Hugging Face at ibm-research/ClariEdge.

6.0

2026-10-23

RESEARCH · 1 SOURCE · arXiv cs.AI

GRAM: Group-wise Rank-Aware Model Merging (arXiv:2610.29110v1)

This paper introduces GRAM, a method that groups geometrically compatible task updates using principal-angle geometry and then applies an SVD-based merge within each group using only fine-tuned weights. GRAM reduces a failure mode called 'subspace interference' and outperforms prior model-merging techniques on benchmarks including GLUE with Flan-T5, an 8-task CLIP vision suite, and medical instruction-tuned Qwen2-7B-Instruct with LoRAs.

6.0

2026-10-22

COMPANIES · 1 SOURCE · AWS Machine Learning

AWS launches SageMaker AI serverless 'Training Sessions' (GA) for model customization

Amazon announced general availability of Training Sessions for SageMaker AI serverless model customization, letting users run custom post-training loops (SFT, DPO, RLAIF, multi-turn RL, LoRA or full fine-tuning) without managing GPU infrastructure; SageMaker handles GPU allocation, distributed execution, checkpointing and fault recovery and offers a cookbook and agent skills for end-to-end workflows. Customers cited use cases including healthcare claim-denial triage and multi-turn agent training with tools like Amazon Neptune and gpt-oss-20b.

7.0

2026-10-22

COMPANIES · 1 SOURCE · AWS Machine Learning

Amazon SageMaker AI adds training session APIs to run custom training loops on managed GPUs

Amazon has launched SageMaker training session APIs that let developers write custom training loops (SFT, RL, preference optimization, etc.) while SageMaker handles the GPU compute, distribution, checkpointing, and fault recovery. The feature includes three SDK clients in the SageMaker Python SDK, an OpenAI-compatible sampling endpoint, and a published cookbook of SFT and RL recipes (including examples on GSM8K and LoRA workflows).

7.0

2026-10-21

RESEARCH · 1 SOURCE · arXiv cs.AI

TxnQA: benchmark for analytical reasoning and evidence grounding over transaction event sequences (arXiv:2610.27256v1)

TxnQA is a new QA benchmark (arXiv:2610.27256v1) for analytical reasoning and evidence grounding over synthetic, privacy-mitigated transaction event streams. It includes 27 task families (81 templates) plus a 5-scenario fraud-risk reasoning suite; evaluates format-following rate (FFR), conditional content accuracy (CA), and end-to-end correctness (E2E); and finds that LoRA-adapted open models can enforce output schemas but still struggle with arithmetic and evidence attribution — e.g., Gemma-4-31B-IT and Qwen3-8B reach E2E scores of 0.867 and 0.816, MoE checkpoints 0.632–0.713, and on the 500-instance fraud suite verdict accuracy often >80% while joint (verdict+exact evidence) accuracy falls to ≈23%.

7.0

2026-10-20

MODELS · 1 SOURCE · AWS Machine Learning

GPT-OSS-20B fine-tuned for multilingual reasoning on Amazon SageMaker HyperPod

Teams fine-tuned the open-weight reasoning model GPT-OSS-20B on a multilingual chain-of-thought dataset (HuggingFaceH4/Multilingual-Thinking) using supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) on Amazon SageMaker HyperPod with EKS, LoRA, TRL/veRL, vLLM, Ray/KubeRay and Amazon FSx for Lustre. Results: the base model almost always produced English chains-of-thought; SFT corrected most of the language behavior gap, and GRPO provided additional improvements only when the desired behavior had non-zero probability in the base model; a Qwen2.5-7B control experiment supported the same conclusion.

6.0

2026-10-16

MODELS · 1 SOURCE · InfoQ AI, ML & Data Engineering

AWS open-sources Strands Decider 2B and launches Strands Box sandbox for agent execution

AWS released Strands Decider 2B, a 2-billion-parameter decision-focused model (based on Qwen3.5-2B with a pointer head and rank-16 LoRA) and published the model weights, training data, and scripts. It also announced Strands Box (developer preview), a kernel-enforced sandbox using the Dogwood policy language to control agent execution (shell, Python, tools, filesystem, network); AWS reports ~115 ms median decision latency on an RTX 3090 and 72.3% accuracy on JevBench for a v19 checkpoint.

7.0

2026-10-16

RESEARCH · 1 SOURCE · arXiv cs.AI

Fine-tuning Qwen2.5-3B (LoRA) to withdraw actions in synthetic NAQD-Env

This arXiv v1 paper evaluates supervised fine-tuning (LoRA) of Qwen2.5-3B-Instruct in NAQD-Env to teach selective withdrawal (suspending actions when support is lost). SFT checkpoints improve aggregate decision agreement (0.77–0.92) and withdrawal recall (0.46–0.86) compared with instruction-tuned baselines (agreement 0.45–0.63, withdrawal recall ≤0.06), but exhibit rule-specific failures (e.g., an original-curriculum checkpoint wrongly withdrew 26/34 actions with alternative support) and other limitations; the authors present a method to detect such regressions and note experimental constraints that limit causal claims.

6.0

2026-10-15

MODELS · 2 SOURCES · Hugging Face · The Decoder

Black Forest Labs releases FLUX.3 Image Klein with bounding-box layout control in Diffusers

Black Forest Labs published FLUX.3 Image Klein to the Hugging Face Hub and ships open checkpoints that run in Hugging Face Diffusers. The model supports text-to-image and multi-reference image-to-image generation, adds bounding-box layout control for region placement and image edits, includes a distilled checkpoint (1024×1024 ~2s on an H100, 16 steps by default), quantized single-file transformer weights (bf16, fp8r, nvfp4), and LoRA/DreamBooth training scripts.

8.0

2026-10-09

RESEARCH · 1 SOURCE · arXiv cs.AI

Harness Compilation (arXiv:2610.11231v1): allocate runtime decisions between frozen small VLMs and an external harness

This arXiv preprint introduces Harness Compilation (HC), an offline procedure where a large teacher rewrites reusable content and control from student execution traces to produce an external harness for a frozen small vision-language model; deployment requires no weight updates or teacher calls. Evaluated on seven VQA settings with students up to 9B parameters, HC reportedly improves scores over bare students by 9.9–23.9 points (averaged across three builds per setting), analyzes five runtime decision types, finds that requesting evidence and open query generation are costly while bounded choices and reading supplied text remain useful, and shows fact cards help all students and that HC with 100 practice items can outperform answer-only LoRA on three tasks, with mixed transfer of decision policies and benefits from recompilation or combining HC with training.

7.0

2026-10-05

RESEARCH · 1 SOURCE · arXiv cs.AI

THPL: Vision-to-language framework for rainbow trout feeding decisions in RAS

This arXiv preprint (arXiv:2610.02378v1) introduces THPL, a vision-to-language decision-support framework for precision feeding of rainbow trout in Recirculating Aquaculture Systems. The method extracts trajectories with Fishsort to compute an Activity Coefficient (AC), encodes temporal and group dynamics with a Hierarchical Behavior Encoder using Temporal and Set Transformers to produce dual 'physical' and 'soft' tokens, and fine-tunes an LLM with LoRA plus counterfactual multimodal Direct Preference Optimization (mDPO); reported results show strong correlation of AC with expert feeding intensity (Spearman ρ = 0.925, p < 0.001) and decision accuracy improvements from 33.33% (text-only) to 93.33% (with dual-evidence tokens) and to 96.67% after mDPO, along with gains in METEOR and diversity metrics.

6.0

2026-10-04

MODELS · 1 SOURCE · The Decoder

NASA and IBM release open-source Lunar Foundation Model trained on 17 years of LRO data

NASA and IBM Research, with academic partners, published the open-source NASA–IBM Lunar Foundation Model, a multimodal foundation model trained from scratch on SomBench — a co-registered corpus of nearly 2 million lunar tile bundles spanning 11 modalities and multiple missions including 17 years of LRO data. The pretrained model improves tasks such as polar ice prediction (up to ~22% error reduction) and coarse-scale crater detection (~19% improvement versus SwinV2-B), and is intended to let researchers adapt lunar observations to downstream tasks with far fewer labels.

8.0

2026-09-30

RESEARCH · 1 SOURCE · arXiv cs.AI

DASA: synthetic continuous embeddings enable effective LLM fine-tuning (arXiv:2609.35868v1)

The paper introduces Desired-Update-Aligned Synthetic Data (DASA), which optimizes continuous synthetic input embeddings via activation-gradient feedback from a frozen reference model and uses them directly for downstream fine-tuning. Experiments on six Llama and Qwen models (1B–32B) across six benchmarks show DASA matches or exceeds natural-language fine-tuning in multiple settings, outperforms GRADMM in most comparisons, and yields a 3.6–4.9× speedup over GRADMM with comparable peak GPU memory.

7.0

2026-09-25

COMPANIES · 1 SOURCE · AWS Machine Learning

Run SkyRL GRPO post-training of Qwen3‑VL‑8B on Amazon SageMaker HyperPod

Amazon demonstrates running the open-source SkyRL framework on SageMaker HyperPod (EKS) to post-train a Qwen3‑VL‑8B vision‑language model with Group Relative Policy Optimization (GRPO). Using a HyperPod Ray cluster with shared FSx storage, the workflow improved maze solve rate from 43.75% to over 95% on a fixed 64‑maze evaluation set and uses colocated vLLM inference and FSDP-sharded policy training with LoRA adapters.

6.0

2026-09-25

RESEARCH · 1 SOURCE · arXiv cs.AI

TW3Cast: frozen router of lightly fine-tuned foundation models ranks #3 on GIFT‑Eval; artifacts released

TW3Cast is a time-series forecasting system that reached rank 3 of 130 on the GIFT‑Eval benchmark by mean MASE as of 2026‑09‑14. It uses no agents or language models: a routing table computed once on the training split (and then frozen) selects among four modes per dataset/configuration — specialists (LoRA or full fine-tunes of Chronos‑2, TiRex or Toto), quantile blends, blends of base models, or a selection tournament run on a backtest — with every decision taken on that training-split backtest. Three guard mechanisms (accuracy+calibration criterion, asymmetric margin for candidates that saw the series in training, and conservative per-window gates) limit selection bias. Reported mean MASE ranks: best base model 33.8, tournament 38.0, full router 19.4. The paper (arXiv:2609.28506v1) releases the routing table, expert index, pinned base-model revisions, score file and a dated snapshot of public scores plus a script to regenerate the leaderboard numbers.

7.0

2026-09-21

MODELS · 1 SOURCE · arXiv cs.AI

LoRA adaptation of DINOv3 ViT greatly improves SAS ATR; later refinements give no further gain

The paper adapts a DINOv3 Vision Transformer to underwater synthetic aperture sonar (SAS) automatic target recognition using a three-stage, parameter-efficient pipeline: Stage 1 applies Low-Rank Adaptation (LoRA) with a frozen backbone, Stage 2 adds hard-negative mining, and Stage 3 uses Supervised Contrastive Learning (SupCon). On at-sea SAS data with a mission-level geographic split and comparisons at 85% test recall repeated over three seeds, LoRA (rank 4) raised AUPRC from 0.300 to 0.679 ± 0.027 while training only 0.26% of weights; the hard-negative and SupCon stages showed no meaningful improvement versus matched controls.

6.0

2026-09-18

COMPANIES · 1 SOURCE · AWS Machine Learning

Amazon SageMaker launches HyperPod Inference Gateway for GPU-aware LLM routing

Amazon announced the SageMaker HyperPod Inference Gateway, a Kubernetes-native EKS addon that routes OpenAI-compatible inference requests using real-time GPU signals (KV cache, queue depth, LoRA residency, etc.) to reduce first-token latency and GPU waste without application changes. The two-tier system offers per-cluster intelligent routing and fleet-wide coordination, deployable via a single InferenceGatewayConfig resource and emitting Prometheus/CloudWatch metrics.

7.0

2026-09-17

RESEARCH · 1 SOURCE · arXiv cs.AI

Fine-tuned Stable Diffusion XL and LLaMA enable controllable Ulos motif generation

arXiv:2609.17987v1 presents a multimodal generative framework that fine-tunes Stable Diffusion XL v1.0 via LoRA and integrates a multimodal LLaMA 1.5-7B to generate culturally faithful Batak Ulos motifs. The system uses four conditioning mechanisms (text, image, representation, semantic map/ControlNet); an ablation across three scenarios found Text+Image+Semantic Map achieved the best FID (270) and CLIP (0.65–0.70) but lowest SSIM (0.65), Text+Image+Representation gave the best overall balance (SSIM 0.84, FID 280), and combining all four worsened FID to 330; qualitative evaluation with nine weavers and 30 public participants showed statistically significant positive acceptance (Wilcoxon p=0.007 and p<0.001), and a web prototype for text- and image-to-image generation was developed.

5.0