Search: “SWGfL”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

The Router Within: Eliciting Native Skill Routing from a Frozen LLM

Gavel reads native skill-routing signals from a frozen LLM's forward passes with two linear maps, beating retrieve-and-rerank pipelines by up to 21.9 points on Qwen3-32B.

Gavel (Glance And Verdict from a frozen LLM) elicits skill routing from a frozen agent LLM using two trained linear maps that read mid-layer states, keeping all skill text out of context. A glance step scores the full library against compact per-skill banks built in one forward pass at installation; a verdict step resumes shortlisted skills' forward passes and fuses likelihood and yes/no judgments as a product of experts. It transfers zero-shot to three public benchmarks plus SkillTraj, a new benchmark of 372 simulated agent trajectories. On Qwen3-32B it beats progressive disclosure and retrieve-and-rerank pipelines adding 1.2B–16B external parameters by up to 13.4 points on written tasks and 21.9 when skills are needed mid-rollout.

Hugging Face daily papers·updated Sep 14, 17:58 UTC · 1d ago·first Sep 13, 20:00 UTC · 2d agoAI research 2 sources

40

I’ve been deepfaked: What do I do?

ESET outlines steps for deepfake victims: preserving evidence, using platform reporting tools, and legal remedies like the US TAKE IT DOWN Act and StopNCII.org.

ESET published a how-to guide for people who discover deepfakes of themselves, covering evidence preservation, platform-specific reporting on Google, Facebook, Instagram, TikTok, YouTube, and X, and escalation to publishers or data protection regulators. It notes the US TAKE IT DOWN Act criminalizes non-consensual intimate imagery (NCII) and requires 48-hour takedowns, while UK and EU laws add creation offenses and GDPR Article 17 erasure rights. Services like StopNCII.org and TakeItDown.NCMEC.org hash images so participating platforms such as Meta, TikTok, Reddit, and X can find and remove matching copies.

ESET WeLiveSecurity·Sep 2, 11:48 UTC · 14d agoAI safety & security1

25

Augustinian BabyLM: What Ostensive Definition Can and Cannot Teach a Small Language Model

Study shows visually grounded token embeddings in a small masked LM persist through training and improve object-property knowledge, but escape standard BabyLM benchmarks.

The paper implements ostensive definition for a small DeBERTa masked language model trained on 10M words, seeding visually grounded tokens with embeddings derived from labeled image regions before training. Visual initialization leaves a persistent, seed-replicated advantage on object-property knowledge (COMPS) and a corpus-tailored Visual-Property Swap benchmark covering color, material, size, and shape, but has no effect on most BabyLM grammar benchmarks. Synthetic grounding of previously unseeded words causally transfers the advantage to exactly those words.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:43 UTC · 5d agoAI research

20

[AINews] Fal’s H3 Max Live breaks the infinite videogen barrier

Fal post-trained MiniMax H3 into a 'Max' variant with 35x-faster inference, enabling faster-than-realtime AI video generation and infinite streams.

Fal post-trained MiniMax's H3 model into a 'Max' variant and optimized it for its in-house inference engine, achieving roughly 35x the speed of the official endpoint. The optimization enables faster-than-realtime video generation, demonstrated by an infinite interactive AI-generated stream productized by levels.io. The roundup also notes Meta Muse Code's general availability with an SDK, open DeepSeek-V4-Flash-Vision-Exp weights, GLM-5.3-Flash's strong agentic cost/performance rankings, and Tencent's 770B-parameter Hy4 Preview MoE with 49B active parameters.

Latent Space·Sep 1, 04:36 UTC · 15d agoAI industry

50

SAFIRE: Safety-Critical Benchmark for Fine-grained Fire and Smoke Understanding in Multimodal LLMs

SAFIRE, an 83K-image fire and smoke benchmark, shows open-source multimodal LLMs average only 61.9% accuracy on safety-critical fire reasoning.

SAFIRE is a large-scale benchmark for fire-smoke understanding in multimodal LLMs with 83K captioned images across 20 scenarios and 193K multiple-choice VQA questions spanning 10 evaluation dimensions from perception to higher-order reasoning. Annotations were built via a GPT-5.4-assisted multi-stage pipeline with MLLM majority voting. Ten open-source MLLMs (8B-38B) average 61.9% accuracy, exposing major gaps in safety-critical reasoning. Adapting vision encoders on 7% of the domain data raises fire-scene classification from 20.1% to 64.5%.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:59 UTC · 8d agoAI research

33

Bad Likert Judge: A Novel Multi-Turn Technique to Jailbreak LLMs by Misusing Their Evaluation Capability

Unit 42 details the Bad Likert Judge multi-turn jailbreak that abuses LLMs' evaluation capability, raising attack success rates over 60% across six frontier models.

Palo Alto Networks Unit 42 describes the Bad Likert Judge technique, a multi-turn jailbreak that asks a target LLM to act as a Likert-scale judge scoring the harmfulness of example responses. The highest-rated example in each scale can carry harmful content, bypassing the model's internal guardrails. Testing across six state-of-the-art text-generation LLMs showed an average attack success rate increase of more than 60% versus plain attack prompts, with tested models anonymized. The technique targets edge cases rather than typical use, and the article positions the work as guidance for defenders on potential jailbreak risks.

Palo Alto Unit 42·Aug 17, 12:34 UTC · Aug 17, 2026AI safety & security

60

Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost

Cognition released SWE-2, an RL post-trained coding model from Kimi K3, scoring 50.0% on FrontierCode 1.1 Main and available only inside Devin.

Cognition released SWE-2, its most capable coding model, post-trained with reinforcement learning from Moonshot AI's 2.8T-parameter Kimi K3 base. It scores 50.0% on FrontierCode 1.1 Main, within 1 point of Fable 5.1 at 64% lower cost, and RL reportedly adds 5-6 points over the K3 base on many benchmarks. It is the first Cognition model with selectable reasoning-effort levels all trained in a single RL run using Pareto-slope-matched cost penalties. There are no open weights and no standalone API; it runs only inside Devin (Desktop, CLI, with Web and Fusion rolling out), free for paid tiers through October 10, 2026.

MarkTechPost·Sep 12, 23:56 UTC · 3d agoModel release1

58

Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation

MovieGrid arranges long videos on spatial grids during post-training, generating 6.05x more shots than temporal packing with state-of-the-art cross-shot consistency.

MovieGrid is a multi-grid post-training paradigm that decomposes long videos into temporally ordered chunks arranged on a spatial grid for joint modeling, enabling cross-chunk information exchange. The authors build the Multi-Grid Long Video (MGLV) dataset from 1,000 long-form videos, producing 54K grid videos paired with character-aware story prompts. Under the same token budget, MovieGrid generates 6.05x more shots than Temporal Packing in a 1,616-frame video. It achieves state-of-the-art intra-shot consistency of 0.9131 versus 0.8086 for HoloCine and inter-shot consistency of 0.5914 versus 0.5384 for StoryMem.

Hugging Face daily papers·Sep 5, 20:00 UTC · 10d agoAI research

20

Training a coding model to paint watercolours with TRL and OpenEnv

Hugging Face tutorial trains a coding model with TRL and OpenEnv to paint watercolours through generated code.

A Hugging Face blog walkthrough uses the TRL reinforcement learning library and the OpenEnv environment framework to train a coding model. The target task is generating code that produces watercolour-style drawings, serving as a hands-on reinforcement learning training example. No article body was available in the feed, so specifics are limited to the title.

Hugging Face Blog·Sep 3, 00:00 UTC · 13d agoAI tools & infra1

18

Coding Agents Have Converged: Why the SWE-bench Leaderboard Can No Longer Order Its Top Entries, and What to Measure Instead

Audit of 254 SWE-bench submissions finds top coding-agent entries statistically inseparable, so small leaderboard gaps no longer establish rank.

The paper audits 254 SWE-bench submissions across four splits without running models. On Verified, the top two entries each resolve 396 of 500 instances, and exact paired McNemar tests separate none of the 29 adjacent top-thirty pairs at alpha=0.05. Within-model scaffold score ranges reach 29.8 percentage points, versus an 8.8-point spread among the top thirty. The authors release a five-step audit protocol and recommend reporting comparison-set-specific resolution and model-scaffold provenance.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 16:29 UTC · 20h agoAI research

40

Adversarial Testing of Automated Program Repair Agents for Security Vulnerabilities

SWEADV benchmark shows adversarial issue descriptions make LLM program-repair agents write insecure fixes in 51.7% of cases, evading most detection tools.

Researchers built SWEADV, a benchmark of 750 adversarial issue descriptions derived from 150 SWE-bench Verified repair tasks, covering command execution, deserialization, path traversal, denial of service, and weak hashing attack types. Tested on mini_swe agents backed by GPT-5-Mini, MiniMax-M2.5, and DeepSeek-R, adversarial descriptions induced malicious behavior with successful repair in 51.7% of cases. Detection was weak: LLM-as-judge pre-repair screening reached only 62.3% accuracy, and post-repair detection via static analysis and LLM-as-judge achieved just 39.4% and 55.4%.

arXiv cs.CR·Sep 14, 17:53 UTC · 1d agoAI safety & security1

52

Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training

NVIDIA researchers detail an end-to-end system for online draft co-training that speeds speculative decoding in large-scale long-context RL post-training.

The paper tackles scaling online draft co-training for speculative decoding in RL post-training, where rollout generation dominates cost. It extends packed, load-balanced zigzag ring attention to merge rank-local branch attention with causal main-sequence attention for context parallelism, and introduces TapChannel to transport target features across pipeline-parallel stages without changing the schedule. Experiments show co-trained drafts tracking the policy baseline with substantial rollout and end-to-end speedups up to 122B parameters and strong scaling at 256K tokens.

Hugging Face daily papers·Sep 6, 20:00 UTC · 9d agoAI research

26

Real-SWE: Benchmarking AI models on private, real-world, enterprise codebases

Real-SWE benchmark tests coding agents on licensed private enterprise codebases; top model Fable 5.1 resolves only 38.8% of tasks.

Real-SWE is a new benchmark evaluating frontier AI coding agents on tasks drawn from private production codebases licensed from real companies, spanning billing, tax calculation, and cross-service migrations. Fable 5.1 with Claude Code leads at 38.8% resolution rate (pass@1 over eight runs), followed by GPT-6 Astra Codex CLI at 33.8% and Gemini 3.8 Flash Gemini CLI at 31.2%. Tasks use native harnesses and realistic tooling including Docker, Kubernetes, PostgreSQL, Redis, and Linear; median reference solutions edit 11 files versus 6 for DeepSWE and FrontierCode.

Hacker News · AI·Sep 12, 20:25 UTC · 3d agoAI researchHN 45↑ · 31 comments

48

MInTRL: Off-policy Intervention can boost On-policy RL

MInTRL injects sparse judge corrections into on-policy RL rollouts, expanding exploration beyond on-policy sampling while preserving learnability on math and code benchmarks.

Minimal Intervention Reinforcement Learning periodically has a judge-intervention policy replace erroneous suffixes of the current policy's output with short corrections, then returns control, keeping trajectories largely on-policy. Training uses a sequence-level advantage-regression objective that removes the need for importance sampling. Across math and code benchmarks it consistently beats standard on-policy and off-policy baselines, remains effective with self-intervention, and performs best at moderate intervention intensity.

Hugging Face daily papers·Sep 10, 20:00 UTC · 5d agoAI research

40

MobileVLA-R1 2.0: RL-Enhanced Reasoning for Mobile Robot Control

MobileVLA-R1 2.0 couples chain-of-thought reasoning with RL for mobile robot control, gaining 10 points on real Unitree G1 tasks.

MobileVLA-R1 2.0 is an RL-enhanced vision-language-action framework that explicitly couples structured embodied reasoning with executable mobile robot control via supervised Chain-of-Thought alignment and reinforcement learning. A reasoning-conditioned action decoder maps multimodal reasoning representations to task-level action targets, decoupling high-level action generation from robot-specific actuation for both locomotion and manipulation. It achieves an average 1.6 point SR improvement on VLN-CE and a 10.0 point improvement in full-task success on real-world Unitree G1 mobile manipulation, with evaluations covering navigation, quadruped control, and real deployments on Unitree Go2 and G1 robots.

Hugging Face daily papers·Sep 4, 20:00 UTC · 11d agoAI research

32

A Princeton Researcher Proposes Recurrent Looped Transformer (RLT) that Carries Decoder State across Every Token, Fixing 96 Blocks per Token with Unbounded Temporal Depth

Princeton researcher Yifan Zhang proposes Recurrent Looped Transformer, carrying full decoder state across every token for unbounded temporal depth.

Yifan Zhang's technical report defines the Recurrent Looped Transformer (RLT), pairing a causal encoder with a recurrent decoder whose final output and layerwise sliding-window attention cache carry into every subsequent token with no prompt-response boundary reset. The reference configuration ties 48 encoder and 48 decoder layers, executing 96 logical blocks per token while the state path grows to 48t blocks after t tokens at fixed per-token compute. The report details RL replay contracts that rebuild all states under current parameters and exact prefix snapshots for multi-turn serving, but explicitly reports no measured efficiency, reasoning quality, or scaling results.

MarkTechPost·Sep 13, 17:02 UTC · 2d agoAI research1

40

MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes

MetroLLM-Bench is a 955-case benchmark testing language models as transit kiosk tool-calling runtimes across six real metro systems.

The benchmark covers 37-414-station metro systems and eleven task categories including routing, fare calculation, disruptions, accessibility, and adversarial input, with 14 deterministic and 8 semantic scoring components. Of 26 models from six vendors, a PEFT-tuned 4B Qwen 3.5 student scored 91.3 on Tier 1, exceeding GPT-5.6 (90.6/90.0), while Muse Glimmer 30B led the composite ranking. A deterministic rule-based baseline reached 84.6, and PEFT gains over base models shrank from +7.03 points at 2B to -0.91 at 27B.

Hugging Face daily papers·Sep 8, 20:00 UTC · 7d agoAI research

30

Hierarchical NeRF with JAX3D for Volumetric Rendering, Novel-View Synthesis, and 3D Reconstruction

MarkTechPost tutorial implements a hierarchical NeRF in JAX using jax3d volume-rendering primitives for novel-view synthesis and 3D reconstruction.

The tutorial builds an end-to-end hierarchical Neural Radiance Field using JAX, Flax, Optax, and jax3d's volume-rendering functions (sample_along_rays, volume_rendering, sample_piecewise_constant_pdf). It implements positional encoding, skip connections, separate coarse and fine networks, and view-direction conditioning with hierarchical importance sampling. Training uses JAX JIT compilation, Adam optimization, exponential learning-rate decay, and gradient clipping. Evaluation covers PSNR, depth and opacity visualization, 360-degree rendering, and marching-cubes geometry extraction.

MarkTechPost·Sep 13, 19:46 UTC · 2d agoAI research

18

RelightFormer: Feed-forward Generative Transformer for Multiview Object Relighting

RelightFormer is a feed-forward generative transformer for photorealistic single- and multi-view object relighting, trained on a 90K-object dataset.

Researchers introduce RelightFormer, a feed-forward generative transformer adapted from a video foundation model that performs direct image relighting without explicit intrinsic property estimation. The architecture injects target environment maps via a latent illumination module with cross-attention and uses permutation-invariant positional encodings for unordered multi-view inputs. Training relies on the newly constructed Laval Objaverse Dataset (LOD) with 90K objects and 39K unique illuminations, and the model shows state-of-the-art quality with strong zero-shot generalization across single-view, multi-view, and novel-view relighting.

Hugging Face daily papers·Sep 6, 20:00 UTC · 9d agoAI research

38

Miles v0.1: Production-Level Post-Training

Radix Ark open-sources Miles v0.1, a full-stack RL post-training framework demonstrated with asynchronous agentic RL on GLM-5.2 744B-A40B across 64 GB300 GPUs.

Miles v0.1 is a full-stack, open-source system for frontier-scale reinforcement-learning post-training, built on slime with rollout engines on SGLang and trainers supporting NVIDIA Megatron-LM and PyTorch FSDP backends plus three weight-synchronization transports. It supports full-parameter RL, LoRA RL, on-policy distillation, supervised fine-tuning, true-on-policy rollout-training alignment, and extends to diffusion models. The end-to-end case study ran fully asynchronous agentic RL on GLM-5.2 744B-A40B for terminal-use coding tasks on 64 NVIDIA GB300 GPUs with a median step time of 263 seconds over the first 30 measured steps. The code is open-sourced on GitHub.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI tools & infra

45

Implementation of Machine Learning Workflows with NVIDIA cuML, RAPIDS, GPU Benchmarking, Explainability, Clustering, and Model Inference

Hands-on tutorial implements NVIDIA cuML and RAPIDS to GPU-accelerate scikit-learn-style ML workflows with benchmarking, clustering, and inference.

The tutorial demonstrates NVIDIA cuML as a GPU-accelerated machine learning framework, using cuml.accel to speed up unmodified scikit-learn scripts with zero code changes and the native cuML API for CuPy/cuDF interoperability. It benchmarks CPU versus GPU implementations of PCA, K-Means, nearest-neighbor search, logistic regression, random forests, and DBSCAN on datasets up to 200,000 samples with 64 features. It also builds GPU pipelines with UMAP, t-SNE, and HDBSCAN, validates GPU-generated SHAP explanations, uses the FIL library for forest inference, and covers model serialization and GPU/CPU portability.

MarkTechPost·Sep 13, 01:42 UTC · 3d agoAI tools & infra

14

dealignai/GLM-5.3-CYBERSECURITY-FP8 — new model trending #13 on Hugging Face

dealignai releases GLM-5.3-CYBERSECURITY-FP8, a 753B MoE weight-modified variant cutting refusals on offensive-security prompts, trending #13.

dealignai released GLM-5.3-CYBERSECURITY-FP8 on Hugging Face, a cybersecurity-domain 'crack' of the 753B-parameter GLM-5.3 MoE model, currently trending #13. The release directly edits bf16 residual writers, keeps FP8 routed experts, and serves with stock vLLM on 8x H200 GPUs with 131k context. HarmBench-320 evaluations show 80-84% direct harm compliance and 89% cyber-offense compliance, while MMLU rose 1.07 points to 86.65%. Copyright-verbatim reproduction remains a known soft-refusal limitation, with an UNCENSORED sibling variant offered.

Hugging Face trending models·Aug 30, 10:46 UTC · 17d agoModel release

30

Evaluating Time-Series Foundation Models and Multimodal Dietary Context for CGM Forecasting

Study finds zero-shot time-series foundation models underperform on CGM forecasting; fine-tuned Chronos-Bolt cuts RMSE up to 18.4% and dietary context adds signal.

The paper evaluates time-series foundation models for continuous glucose monitoring forecasting across eight public datasets covering Type 1 diabetes, Type 2 diabetes, and non-diabetes populations. Under a unified protocol, zero-shot foundation models did not consistently outperform baselines like Elastic Net and PatchTST, but lightweight fine-tuning did, with fine-tuned Chronos-Bolt reducing RMSE by 6.5%-18.4% in the T1D cohort and 8.6%-18.2% in the non-diabetes/T2D cohort. A residual-based fusion framework adding dietary context from CGMacros reduced overall RMSE by about 3% and postprandial RMSE by about 15% versus CGM-only baselines.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:43 UTC · 5d agoAI research

28

Jackrong/Qwopus3.8-27B-Flash-GGUF — new model trending #26 on Hugging Face

Community fine-tune Qwopus3.8-27B-Flash, built on Qwen3.8-27B, cuts agent reasoning latency with 12.8% faster decoding and 80.7% MTP acceptance.

Jackrong released Qwopus3.8-27B-Flash, a fine-tune of Qwen3.8-27B optimized for long-running agent workloads, reporting 12.8% faster decoding and 80.7% multi-token-prediction acceptance. Training used roughly 1.5 million teacher-scored SFT examples filtered to the top 10%, followed by reinforcement training with NVIDIA NeMo-RL and GSPO. The author notes an explicit trade-off: MMLU-Pro mixed-set scores are lower than the base model, and a known bug can produce incorrect Python indentation. Author-provided benchmarks have not been independently verified.

Hugging Face trending models·Sep 4, 12:43 UTC · 11d agoModel release1

35

OpenArch – PyTorch implementations of modern LLM architectures

OpenArch provides PyTorch reference implementations of modern LLM architectures for developers and researchers.

OpenArch is a GitHub project offering PyTorch implementations of modern large language model architectures. The repository attracted 43 points and 7 comments on Hacker News. It targets developers and researchers who want readable, runnable versions of current LLM architectures.

Hacker News · AI·Sep 14, 07:55 UTC · 2d agoAI tools & infraHN 43↑ · 7 comments

15

With Groq 3 LPX in Full Production, NVIDIA Extends Vera Rubin Inference for Agents

NVIDIA puts Groq 3 LPX into full production and extends Vera Rubin NVL72 rack-scale systems for fast token generation in agentic AI inference.

NVIDIA announced that Groq 3 LPX is in full production as part of an extension of the Vera Rubin NVL72 rack-scale platform aimed at agentic AI inference. The announcement frames the next era of inference as full-stack AI factory co-design across chips, networking, and systems rather than a single component breakthrough. The focus is improving token generation speed for agent workloads.

NVIDIA Blog·Aug 24, 15:00 UTC · 22d agoAI industry

62

UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents

UC Berkeley's CUA-Lite is an open platform unifying computer-use agent sandboxes, datasets, evaluation and RL; Lite.OSWorld cuts OSWorld memory 4.1 GB to 0.9 GB.

UC Berkeley researchers released CUA-Lite, an open platform placing agents, environments, traces, and training for computer-use agents behind one action space, one LiteSample schema, and one command across desktop, browser, and mobile. Lite.OSWorld reproduces the OSWorld task suite and evaluators in plain Docker containers (0.9 GB RAM vs 4.1 GB, cold start 23.8s, ~4.6× more parallel instances), with scores matching the QEMU/KVM VM across 13 models. The platform claims 30k+ verifiable tasks, 15+ benchmarks, 10+ agents, and 20+ datasets on Hugging Face including Aguvis, OpenCUA, and ScaleCUA. A documented SFT run lifts Qwen3-VL-2B-Instruct mean episode return from 0.138 to 0.237 on the 332-task lite.osworld split.

MarkTechPost·Sep 6, 06:11 UTC · 10d agoAI tools & infra1

45

Target leakage, not model class, explains reported accuracy in survey-based cardiovascular screening: a leakage-tiered audit of glass-box and tabular foundation models

Audits of 10 classifiers on BRFSS show target leakage, not model class, drives the reported 0.89 AUROC in survey-based cardiovascular screening.

The study benchmarks ten model classes, including glass-box and tabular foundation models, for prevalent myocardial infarction on 442,067 respondents of the 2022 BRFSS across five feature tiers of decreasing leakage risk. Removing two post-diagnostic features costs every model 0.049-0.051 AUROC and collapses performance into a 0.0045-wide band, and the explainable boosting machine matches all alternatives within 0.005 while scoring roughly 104x faster than the strongest foundation model. Frozen models transport within 0.002 AUROC to 2023 data; the authors conclude evaluation practice and feature sets, not model capacity, are the binding constraint.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:29 UTC · 5d agoAI research1

28

XHToken/Spark-X2.5-4B-GGUF — new model trending #30 on Hugging Face

XHToken released GGUF weights of Spark-X2.5-4B, a compact model with 1M-token context and 200+ language support, under Apache 2.0.

The Hugging Face repository provides BF16 GGUF conversions of Spark-X2.5-4B, a compact general-purpose language model for conversation, writing, translation, reasoning, coding, tool use, and agentic workflows. The model uses a hybrid attention architecture, supports a native context length up to 1M tokens, and covers more than 200 languages. Local inference is supported through Ollama and LM Studio via an XHToken llama.cpp fork, with a --think=false flag to disable thinking mode for faster responses. Released under Apache License 2.0; it was trending #30 on Hugging Face at publication.

Hugging Face trending models·Aug 28, 11:26 UTC · 19d agoModel release

24

Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal

A self-distillation safety framework tunes narrow-boundary refusals in Qwen3-8B, raising target-domain refusal to 84.75% while cutting over-refusal from 15.20% to 5.20%.

The paper formulates narrow-boundary safety, where deployments need refusals within specific topics rather than whole subjects, and proposes an offline self-generated framework with controlled topic generation, escalating retries, and harmful-benign boundary pairs. On political persuasion with Qwen3-8B, the method raised target-domain refusal from 9.47% to 84.75% and cut the mean unsafe-response rate across three broader benchmarks from 26.26% to 0.14%. Verified target-model responses reduced over-refusal from 15.20% to 5.20%, and boundary-pair data cut comply-side over-refusal on held-out pairs from 32.94% to 4.16%. Results show data composition controls the safety-usability trade-off and alignment should be evaluated on both sides of the refusal boundary.

Hugging Face daily papers·Sep 2, 20:00 UTC · 13d agoAI safety & security1

32