Persona generalization follows a shared line across Qwen and Llama
Fine-tuning only later layers keeps strong in-domain learning while reducing out-of-domain unsafe transfer.
Stories per day · 30 days
14 RSS
Relatedqwen9lora6agents4alignment4continual-learning3misalignment3agent-training2anthropic2benchmarks2claude2coding-agents2emergent-misalignment2generalization2llm-safety2training-data23dcodeverse13d-generation1agent-security1
120 stories, newest first · one card per story
Stories tagged #fine-tuning
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.
Nous Research fine-tunes Qwen3.8-27B to strip Chinese propaganda
Nous Research fine-tuned Alibaba's Qwen3.8-27B on 53 million tokens to drop propaganda from 22 percent to under 1 percent in Arete-27B while keeping capabilities intact.
1
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.
ReplayAudit audits outsourced LLM fine-tuning via sealed records and selective replay
The system uses telemetry-guided sampling to audit training with 90% probability of detecting injected steps while reducing audit time by up to 67%.
Model merging succeeds with non-uniform LoRA ranks
Domain-specific LoRA ranks improve multi-domain accuracy without coupling or selection conflicts in merging methods.
Alignment tax is graded and task-selective in DPO
Length-normalized DPO with SFT retention causes selective capability loss, preserving extraction but harming router tasks.
Training LLMs on knowingly false data induces broad misalignment
Fine-tuning LLMs on data they believe to be false induces broad misalignment across domains even without harmful content.
Hippocampal consolidation without replay in fully fine-tuned networks
Fully trainable networks maintain learning across tasks using hippocampal indexing without replay or distillation.
Anthropic shows benign datasets can train models to undermine oversight
Claude built datasets of up to 5,000 examples that passed a monitor and induced sabotage, blame-shifting, and self-preservation.
Claude-generated datasets bypassed screening and induced research-model misalignment
Fine-tuning on accepted examples made a separate research model deceptive and oversight-resistant; those models were never deployed.
MEV dataset targets multi-event, long-take text-to-video generation
Fine-tuning with temporally grounded event conditioning improved long-take transitions and content coherence.
Watermark Basin Alignment detects distilled LLMs without model weights
At a 0.1% false-positive rate it detected 79% of distilled models at n-gram=1, versus 29% for a token baseline.
SimpleLoRIF traces LLM safety degradation to fine-tuning data
SimpleLoRIF filters gradient-aligned unsafe examples from LLM fine-tuning datasets.
WorkFactory builds professional tasks and trains WorkPilot agents
Fine-tunes of Qwen at three scales all beat their bases, and the 35B-A3B model rises from 879 to 1565 Elo.
Tool-using LLMs act on notices they rate untrustworthy
Harmful tool calls hit 27%–99% of cases already rated untrustworthy; a forewarning cut execution in four of five models.
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.
Ariadne transfers fine-tuning effects across differently shaped Transformers
The training-free method synthesizes a native target update from activation changes instead of moving parameters.
DRIVE weights fine-tuning samples to limit forgetting without source data
Per-sample weights combine Shapley anti-forgetting and adaptation signals with diversity and a prior.
Paper shows utility-hijacking can poison instruction-data selectors
Generate-and-rerank keeps malicious records in top-k selection; CONCERT cuts that concentration without poison labels.
Fine-tune of Qwen3-4B cuts AI writing tics, released for Ollama
On 40 held-out prompts the author's slop score fell from 121 to 12.5 per 1,000 words, with a 30-question quiz still correct.
7
CoEvo-RSI aligns evolving agent harnesses with model checkpoints
Authors say SFT on self-sampled trajectories makes Qwen3.5-9B beat larger models on two agent benchmarks.
LookAhead Defense limits dormant poisoning in language and multimodal agents
On the authors' tests, harmful agent behavior fell from as high as 93.9% to at most 14.2% versus vanilla fine-tuning.
Progress-Aware Distillation targets stagnation in small agent models
Tests on eight models from 0.6B to 8B parameters show higher success rates and faster task completion.