LoRA subspace projection suppresses backdoors but does not erase them
Few-shot reactivation restored 91.7% to 100% attack success in one epoch for every weight-editing defense tested.
Stories per day · 30 days
7 RSS
Relatedfine-tuning6continual-learning5video-generation5minimax4open-weights4huggingface3image-generation3qwen3algonauts2brain-encoding2comfyui2distillation2model-merging2peft2robotics2supply-chain24d-perception1abliteration1
86 stories, newest first · one card per story
Stories tagged #lora
VeilTune fine-tunes LoRA adapters entirely inside secure MPC
From one 32-sample mini-batch of hidden states, authors copied a provider adapter on 99.2% of test inputs.
Telefónica submits ASR system for Speech-COSER Challenge 2026
An ensemble of seven LoRA systems reached 22.17% normalized word error rate on elderly rural Spanish speech.
LoRA merge audit finds bad adapters inflate forgetting scores
Naive merging of an audited four-adapter set cut GSM8K by 10.5 points and HumanEval by 18.3.
Deletion Tuning matches full fine-tuning by zeroing under 0.05% of weights
A binary mask on gate projections stays near base prior-task accuracy while FFT, LoRA, and random sparse tuning lose up to 5.4 points.
LoRA dynamics differ between supervised fine-tuning and reinforcement learning
RL concentrates LoRA spectrum while SFT expands it at moderate rates, with group-mean baseline subtraction identified as the source.
Model merging succeeds with non-uniform LoRA ranks
Domain-specific LoRA ranks improve multi-domain accuracy without coupling or selection conflicts in merging methods.
Intel TDX enables bilateral protection for private LoRA inference
Intel TDX system protects client inputs, outputs, and proprietary LoRA parameters during cloud inference via remote attestation and verified byte consumption.
BISECT improves low-rank continual learning for Vision Transformers
BISECT partitions rank budget into invariant carrier and novelty complement, outperforming baselines with no extra parameters.
Surprise-driven continual learning absorbs planted lies about known facts
A model's own prior separates known-fact lies at AUC 0.78–0.83 but not legitimate updates, at AUC 0.53.
Intrinsic Distribution Matching improves few-step image distillation
At two network evaluations, IDM scored 94.25% on Chinese and 90.20% on English LongText-Bench.
Black-box signatures proposed to detect LoRA adaptations of diffusion models
Tests use only the intrinsic dimension of generated images and assume unknown adaptations and model families.
PREEN predicts editor presets for exemplar color transfer
The model regresses 29 smartphone-editor sliders, runs at sub-second on-device latency, and leads PST50 fidelity metrics.
Venus-DefakerOne2 reports new scores across image, video and spoof forensics
Authors report 96.7% on DeepFake faces and 96.0% on AI-generated videos after training on tens of millions of samples.
Open-H3-Regenerate-2K upscales MiniMax-H3 drafts to 2560×1440
Authors report a five-second clip falling from 87.26 seconds to 22.45 seconds on one B200 with the fast decoder.
3
Rollback backtracks world models to improve robot policies
Absolute success gains were 10.67%, 11.11%, and 11.85% on three real-world tasks versus the strongest baseline.
RoboJev-27B open model scores robot manipulation success from video
On 268 held-out SO-101 episodes the authors report 95.2% balanced accuracy and 0.993 AUROC.
TopKLoRA sparse adapters localize and ablate sleeper-agent backdoors
On three small models, ablating a TopKLoRA circuit removes a planted backdoor while instruction-following holds, with much smaller circuits than dense LoRA.
Licon VBVR LoRA adapts MiniMax-H3 for image-to-video
The adapter, H3-VBVR-I2V.safetensors, is trained on rebuilt VBVR data and loads in a MiniMax-H3 ComfyUI workflow.
AnyGameRenderer restyles game footage while keeping entities
A LoRA on 33B MiniMax-H3 raises temporal entity match from 50.9% to 62.9% on 59 held-out cases.
Poisoned open weights can hide triggers that steal agent credentials
A Qwen2.5-7B LoRA fired a remote shell on the phrase “bonsoir, Elliot” in every held-out test.
2
SolarRenderer adapts MiniMax-H3 for four-step structure-conditioned video
Two LoRAs keep a frozen 33B backbone and render about 12.3 times faster than 50-step sampling.
WeAct selects LoRA adapters as actions for long-horizon agents
On Qwen3-32B, WeAct beats full fine-tuning by 6.2% and static LoRA merging by 27.5% on eight deep-research benchmarks.
BrainLoRA adapts backbones up to 104B parameters for brain encoding
An ensemble of more than 200 LoRA adapters ranks first on the public Algonauts post-challenge leaderboard.