Search: “preventative-steering”

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

Active Adaptation, Not Static Defense: Temporal Dynamics of Preventative Steering in Adversarial Fine-Tuning

Researchers analyze why Preventative Steering protects LLMs against malicious fine-tuning, finding active adaptation drives protection, and propose Progressive Intensity Scheduling.

The paper studies Preventative Steering, a training-time defense that injects undesirable-trait persona vectors during adversarial fine-tuning and removes them at evaluation time. Temporal analysis shows protection emerges from an early compensatory adaptation phase followed by a steady-state phase, with attention output projections acting as the dominant residual-write route for defensive updates. Intervention Delta Preservation experiments show that preserving or reinjecting weight offsets fails to maintain protection, indicating reliance on active adaptation rather than a static defense. The proposed Progressive Intensity Scheduling improves safety robustness on Qwen2.5 and Gemma-3 while reducing harmful trait expression.

arXiv cs.CR·Sep 9, 13:19 UTC · 7d agoAI safety & security1

32

Show-Harness: Just a VLM Agent Can Play Robots

Show-Harness lets VLM agents control robots via discrete semantic action units, outperforming VLA baselines zero-shot and after light fine-tuning.

Show-Harness is an embodied agent harness that exposes discrete semantic action units a VLM reasons over, with embodiment-specific interpreters grounding them into local robot actions. It enables zero-shot robot control with closed-source frontier VLMs and low-cost adaptation of small open-source VLMs using only a few GPU-hours of fine-tuning. The companion GUMI (GUI Manipulation Interface) extends the same semantic action space to GUI-based demonstration collection without specialized teleoperation hardware. Experiments show robust generalization across tasks, embodiments, and environments, beating representative agentic and VLA paradigms.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 17:53 UTC · 6d agoAI research

32

EVOHARNESSBENCH: Can Your Agents Keep Pace with an Evolving Harness?

Researchers introduce EVOHARNESSBENCH, a benchmark showing that evolving agent harnesses (tools, skills, agents) cause forgetting and inconsistent adaptation across 802 tasks.

The paper introduces EVOHARNESSBENCH, a benchmark that places non-stationarity in the externally supplied agent harness rather than in the task stream, evaluating agents across tools, skills, and specialist agents. It comprises 17 multi-stage harness streams built deterministically from verifier-based benchmarks, totaling 802 tasks, 520 tools, 42 skills, and 62 agents. Evaluation covers deployment (retention of previously accessible competence) and self-evolving adaptation settings. Results show harness expansion alone degrades previously solved tasks (harness-induced forgetting), adaptation gains are inconsistent, and retention and adaptation can pull in opposite directions.

Hugging Face daily papers·Sep 2, 20:00 UTC · 13d agoAI research

40

Perturbation Probing: A New Diagnostic for the Fragility of LLM Safety

Unit 42 research shows LLM safety refusals concentrate in a thin neural layer, motivating external, multi-layered AI security controls.

Palo Alto Networks Unit 42 introduces Perturbation Probing, a diagnostic technique for measuring the fragility of LLM safety mechanisms. The research finds that safety refusal behavior is localized within a thin neural layer, implying small perturbations can undermine built-in refusals. The authors argue this motivates external, multi-layered security defenses on top of model-internal safety training.

Palo Alto Unit 42·Aug 28, 22:00 UTC · 18d agoAI safety & security

45

ENCP: Episode-Normalized Conformal Prediction for Vision-and-Language Navigation

ENCP calibrates conformal prediction per navigation episode, giving step-level coverage guarantees for vision-language navigation agents despite within-episode dependence.

Episode-Normalized Conformal Prediction (ENCP) rescales a nonconformity score by a VLN policy's residual confidence and calibrates one maximum score per episode, preserving step-level coverage of at least 1−α despite dependence among steps within an episode. Across four VLN policies and three nonconformity scores on R2R and REVERIE, ENCP meets all reported empirical step-coverage targets in seen-to-unseen evaluation. The model-agnostic uncertainty estimates can signal when an agent should defer to a stronger predictor or human assistance.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:42 UTC · 22h agoAI research

12

The Missing Boundary: How Autonomous Agents Lose Control

Tencent research finds agents lose control in 55-62% of trajectories when degraded control boundaries coincide with executable unsafe opportunities across five models and 16 domains.

The study independently manipulates goal pressure, control degradation, and executable unsafe opportunity in a deterministic multi-turn environment across five agent models and 16 operational domains. Neither factor alone causes substantial loss of control; when both are present, loss-of-control rates reach 55% in the full-factorial study and 62% across ten additional domains. Restoring the original control boundary reduces the rate to 0% even when unsafe actions remain executable, and a context-management ablation shows compaction is harmless when constraints are preserved but omission raises the rate to 87%.

arXiv cs.CR·Sep 10, 03:02 UTC · 6d agoAI safety & security2

50

Qwen-Drive 1.0 tells you why it brakes, just don't expect the explanation to match the maneuver

Alibaba's Qwen-Drive 1.0 adds 3D perception and planning modules to Qwen3.5-4B for driving tasks, though explanations often mismatch maneuvers.

Qwen-Drive 1.0, built on Qwen3.5-4B, combines spatial perception, traffic question answering, and route planning in one vision-language model, adding a bird's-eye-view perception module and a Planning Expert trained via staged fine-tuning and reinforcement learning. The paper finds text-image models do not inherently grasp 3D space; spatial accuracy only improved when the base vision-language model itself was trained on spatial tasks, while avoiding catastrophic forgetting of general knowledge. The cut reinforcement learning-trained version halved road-departure rate in simulation from 24% to 12%, and the model beats specialized driving models in most of Qwen's benchmarks, but its explanations sometimes conflate causes like distant red lights and crossing children, and results partly rest on self-designed tests. The work follows prior findings from PaLM-E and a UC Santa Cruz adversarial sign attack on DriveLM showing VLM driving models' reasoning and spatial gaps.

The Decoder·Sep 7, 12:15 UTC · 9d agoAI research

40

The Evolution of the Agent Harness

Latent Space essay argues late-2025 agent gains came from models and harnesses maturing together, with harness logic absorbed into model weights.

The piece defines the agent harness as everything beyond model weights—tools, context, memory, guardrails—and charts its evolution from ReAct prompting (October 2022) through AutoGPT's premature autonomy, Cursor/Copilot's human-in-the-loop retreat, and Devin's roughly 15% success rate, to o1's capability overhang and Claude Code's February 2025 terminal agent with permission rules. It argues the Christmas 2025 jump cited by Transformer co-inventor Lukasz Kaiser reflected model and harness curves crossing, and that remaining harnesses will serve human attention rather than the model.

Latent Space·Aug 22, 07:30 UTC · 25d agoAI tools & infra

32

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

Researchers introduce SchemeArena, a 400-scenario benchmark stress-testing scheming in LLM agents, finding explicit instrumental goals are the strongest driver of covert misaligned behavior.

The paper presents SchemeArena, a 400-scenario benchmark built through factorized scenario synthesis spanning safety-relevant tool domains, instrumental goals, oversight conditions and pressure mechanisms. The accompanying SCOUT monitor grounds multi-criteria scheming judgments in evidence drawn from agents' reasoning and actions. Stress tests across five LLM agents show explicit instrumental goals are the strongest driver of scheming propensity, while action-only monitoring increased scheming in several closed models, suggesting partial oversight can act as an optimization constraint. The benchmark, code and monitor are released at github.com/launchnlp/SchemeArena.

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

40

Why Patch Automation Needs Brakes, Not Just an Accelerator

Action1's field CTO argues patch automation needs staged deployments and stop conditions, not just speed.

Gene Moody, Field CTO at Action1, writes on BleepingComputer that patch automation must pair acceleration with safeguards. He recommends staged deployment rings with predefined go/no-go criteria, keeping human judgment for domain controllers, databases, and ERP systems. The piece warns that automation without brakes can push a bad update to 10,000 endpoints as fast as a good one.

BleepingComputer·Sep 14, 14:01 UTC · 2d agoIndustry

12

HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness

Researchers introduce HarnessVLN, a zero-shot training-free agent harness that sets new training-free SOTA on vision-language navigation benchmarks including R2R and HM3D.

HarnessVLN is a zero-shot, training-free framework for embodied vision-language navigation that coordinates perception, retrieval, grounding, navigation, recovery, and termination through a unified tool interface. It validates planner proposals against spatial evidence, geometric feasibility, and subgoal consistency, using hierarchical event memory and a persistent Spatiotemporal Graph that stores reusable spatial evidence and failure annotations. It reports success rates of 60.8% on R2R, 53.9% on RxR, 76.0% on HM3D-v2, and 59.3% on HM3D-OVON, surpassing prior training-free state of the art, with real-world humanoid deployment demonstrated.

Hugging Face daily papers·Sep 13, 20:00 UTC · 2d agoAI research

30

DriveZero: End-to-End Driving Beyond Human Demonstrations

DriveZero pairs a frozen vision-foundation-model perception stack with a PPO-trained closed-loop RL teacher to beat replay experts on nuPlan.

DriveZero is an end-to-end camera-only autonomous-driving planner that separates perception and action. Its DriveVFM perception backbone consolidates frozen vision foundation models (DINOv3, SigLIP2, SAM, Depth Anything V2) from raw images without task annotations, while DriveRL trains a privileged PPO teacher policy through closed-loop rollouts in interactive worlds built from real driving logs. The planner distills this teacher, achieving a 93.57 mean nuPlan score across Val14, Test14-hard and Test14-random splits and beating the Log-Replay expert on all three. It also sets state of the art on NAVSIMv1, NAVSIMv2 and closed-loop HUGSIM without human trajectory supervision.

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

30

Context Engineering Inside the Harness: 4 Mechanisms That Beat Context Overflow and Goal Loss on Long-Horizon Tasks

Survey of four harness mechanisms—context budgeting, compaction, todo-state, and memory—that keep long-horizon LLM agents on task across 200+ tool calls.

The article details how agent harnesses, not larger context windows, solve context overflow and goal loss on long-horizon tasks, citing Chroma's Context Rot report showing 18 LLMs (GPT-4.1, Claude 4, Gemini 2.5, Qwen3) degrade on long inputs. Concrete implementations include LangChain Deep Agents offloading tool responses over 20,000 tokens to the filesystem and truncating old tool calls at 85% window usage, and Claude Code capping auto memory at 25KB while re-reading the 5 most recently modified files after compaction. OpenAI's Responses API now offers server-side compaction via context_management with a standalone /responses/compact endpoint, which Codex uses for long-running coding tasks. Manus reports a roughly 100:1 input-to-output token ratio per ~50-tool-call task, motivating todo.md state recitation to prevent goal drift.

MarkTechPost·Sep 13, 05:56 UTC · 3d agoAI research1

52

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

ASTRIL-MPC: Autonomous Traversal Framework of Articulated Tracked Robots with Language-Guided Neural-Kinematic MPC

ASTRIL-MPC combines learned kinematics, NMPC, and LLM-guided safety-checked retuning for articulated tracked robot traversal in search-and-rescue.

ASTRIL-MPC is a language-guided neural-kinematic model predictive control framework for autonomous traversal of articulated tracked robots in urban search and rescue. A learned kinematics model predicts short-horizon task-state increments, NMPC plans with feasibility constraints, and an LLM proposes bounded, safety-checked updates to weights and bounds. The compiled predictor enables a full control cycle within 100 ms, improving traversal-quality scores by up to 71% over non-adaptive NMPC and 67% over a PPO baseline while eliminating measurable collision impacts.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:17 UTC · 4d agoAI research

12

Show-Harness: Just a VLM Agent Can Play Robots

Show-Harness enables VLM agents to control robots via a semantic action interface, achieving zero-shot frontier control and few-GPU-hour adaptation of small VLMs.

Show-Harness exposes discrete semantic action units that VLMs reason over, with embodiment-specific interpreters deterministically grounding them into local robot actions. It enables zero-shot closed-source frontier VLM control and adapts small open-source VLMs for low-cost deployment with a few GPU-hours of fine-tuning. The companion GUMI interface extends the same semantic action space to GUI-based demonstration collection without teleoperation hardware, and Show-Harness-equipped agents outperform representative agentic and VLA paradigms.

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

28

Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration

Researchers introduce Decoy Direction Optimization, a cheap weight-editing defense that blinds refusal-direction ablation attacks against open-weight LLM safety guardrails.

Refusal Feature Ablation bypasses safety guardrails in open-weight LLMs by projecting out a linear refusal direction, often with high attack success rates. Decoy Direction Optimization injects a high-magnitude nonlinear decoy into MLP neurons so attackers' contrastive estimators ablate a harmless orthogonal feature instead. Evaluated across six model families, DDO keeps ASR below 10% under standard RFA and on Llama-3-8B-Instruct reduces Heretic weight-level attack ASR from 88.7% to 18%. It costs 30 to 450 times less per configuration than trained defense baselines.

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

55

Steering Geometry: Validating Human Value Geometry in LLM Steering Space

Researchers show LLM activation steering vectors encode human value geometry predicted by Schwartz's theory, with fidelity improving at scale but degrading after instruction tuning.

The study tests whether latent geometry of activation steering vectors reflects theory-specified structure in human values, using Schwartz's Theory of Basic Human Values and a new 26K-sample benchmark spanning 20 values. Distribution-driven methods (CAA, SphericalSteer, ODESteer) recover predicted value topologies with Spearman correlation up to 0.51 (p < 10^-13), while behavior-centric methods (COLD-Steer, BiPO) steer comparably but show little geometric fidelity. Geometric alignment improves with model scale but drops after instruction tuning. Better geometric alignment also yields more human-consistent cross-value transfer, steering one value lifting compatible values and suppressing opposing ones.

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

38

When the World Lies: Backdoor Attacks on Latent World Models for Downstream Control

A poisoned world-model checkpoint hijacks downstream controllers without an explicit trigger rule, passing clean-data evaluation while steering 100% of triggered actions.

Researchers show that a released pretrained world-model checkpoint acts as a supply-chain backdoor for downstream control. The poisoned model routes trigger-bearing observations into a chosen latent region and reshapes dynamics so the victim's own Dreamer-style actor training or MPC/CEM planning re-discovers attacker-targeted actions. The attack hijacks 100% of triggered steps in the strongest settings while retaining roughly 75% clean-task success and passing standard clean-data diagnostics. Moderate clean fine-tuning fails to remove the backdoor without substantially degrading clean control.

arXiv cs.CR·Sep 14, 16:01 UTC · 2d agoAI safety & security

40

PlannerForge: LLM Agents for Scenario-Based Testing of Motion Planners in Autonomous Driving

PlannerForge unifies scenario-based testing of autonomous driving motion planners in one LLM-agent framework, outperforming prior baselines.

PlannerForge is an LLM-agent framework that covers the full scenario-based testing pipeline for autonomous driving systems, spanning scenario generation, selection, modification, routing, planner testing, plus new enhancement and benchmarking stages. In evaluations with 10 off-the-shelf LLMs, best-per-task scores range from 0.88 to 1.00, and open-source 20-35B backends such as Qwen3.6:35B match commercial APIs on most tasks. End-to-end chaining retains 83% (commercial) and 78% (open) of seed queries, beats Scenario Factory 2.0 on executable generation, and cost-tuning lifts planner success from 50.4% to 70.2% while cutting collisions from 19.0% to 8.4%.

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

28

Inoculation Midtraining with Learned Neologisms

Inoculation Midtraining confines unsafe LLM behavior to a neologism-marked context, reducing misalignment after unsafe post-training but leaking under nearby contextual cues.

The paper introduces Inoculation Midtraining, which teaches a base model during midtraining that unsafe behavior belongs to a context marked by a learned neologism token, then post-trains on unsafe data within that context. Across supervised fine-tuning and RL post-training regimes, the technique reduces misalignment while preserving transfer of benign properties like German or Shakespearean prose. However, it does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. The authors conclude it is not yet a load-bearing component of a developer safety framework.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:12 UTC · 1d agoAI safety & security

42

Safe Meta-Reinforcement Learning via Information Space Reachability

Safe meta-RL framework reasons about safety in information space, learning a safety value function used for safety filtering and constrained policy optimization.

The paper proposes safe meta-RL that reasons about safety in information space, capturing both physical state and the agent's belief over the underlying task. A safety value function measures the probability of avoiding unsafe regions indefinitely and satisfies a self-consistency condition and Bellman equation, making it learnable via meta-RL. The resulting algorithm uses the learned function for safety filtering and constrained policy optimization, with effectiveness demonstrated on meta-RL benchmarks.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:27 UTC · 1d agoAI research

25

Can We Stop The Ads? Taxonomy and Characterization of Smartphone Splash Ads and Existing Countermeasures

Study of 108 ad-defense implementations finds only one tool blocked splash-ad navigation across ten popular apps, and it required Accessibility permission.

The paper taxonomizes smartphone splash ads — full-screen ads at app launch that trick users into trigger mechanisms such as moving the phone — and analyzes 108 documented advertising defenses for deployment barriers. Many defenses require device rooting, jailbreaking, runtime code injection, or application modification; others need extra permissions, rule maintenance, compilation, or payment. In evaluating 13 configurations of 11 tools across 10 popular apps, only one prevented ad-triggered navigation across all ten apps, requiring Accessibility permission and leaving ads visible roughly one second before dismissal. Documented harms include delayed emergency response, driver distraction, and degraded accessibility for vision-impaired users.

arXiv cs.CR·Sep 15, 15:24 UTC · 1d agoResearch

32

What Matters, When? Diagnosing and Improving Conditional Visual Grounding in Visuomotor Imitation Policies

Researchers diagnose conditional visual grounding failures in visuomotor imitation policies and show targeted interventions substantially improve distractor robustness.

The paper studies why ACT-based visuomotor imitation policies fail when visually similar distractor objects or receptacles are introduced, finding sensitivity depends on both distractor type and manipulation stage. Interventions including distractor augmentation, phase-dependent attention regularization, and appearance-based visual prompting improve target selection while preserving spatial control information, with gains in simulation and on a physical UR3e. The same failure pattern is confirmed in a pretrained vision-language-action policy on a state-conditioned medical instrument-handling task.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:27 UTC · 11d agoAI research

18

The safety penalty: Reclaiming operational sovereignty in the age of AI

Cisco Talos argues restrictive frontier AI models impose a 'safety penalty' on security teams, urging operational sovereignty for defensive AI in incident response.

Cisco Talos published commentary arguing that increasingly restrictive frontier AI models create a 'safety penalty' that slows real-time incident response. It recommends organizations pursue operational sovereignty so defensive AI can keep pace with unconstrained adversaries.

Cisco Talos·Aug 25, 10:00 UTC · 22d agoIndustry

20

“Sorry, I can’t help with that”: How your guardrails might become the attacker’s best friend

Cisco Talos's David Bianco argues AI guardrail customization requires operational sovereignty so defenders retain the advantage over attackers.

In his first Threat Source newsletter, Cisco Talos's David Bianco explores how AI guardrails could end up aiding attackers and argues that operational sovereignty is needed when customizing them. The piece stresses that organizations should control their own AI safety configurations to keep the defender's advantage. This is commentary and analysis rather than a report of a new incident or vulnerability.

Cisco Talos·Aug 27, 18:00 UTC · 19d agoAI safety & security

22

The Model Proposes, the Code Disposes: A Pre-Registered Ablation of a Verifier-and-Acceptance Stage in an LLM-Orchestrated Offensive-Security Agent

Pre-registered ablation finds a model verifier stage in an LLM offensive-security agent suppresses findings; removing it eliminated suppression with precision tradeoff.

The paper evaluates a verifier-and-acceptance stage in an LLM-orchestrated offensive-security agent via a pre-registered 20-run confirmatory ablation and a 2x2 factorial study with 40 runs on vulnerable lab targets. Removing the stage eliminated pre-report suppression (median 2 vs 0 findings, p = 0.00003) but reduced model-blinded shipped precision (0.471 vs 0.353, p = 0.0087). Suppression was attributed to the model verifier rather than deterministic acceptance rules, and an instrumented canary recorded zero external contacts in all 60 runs. The full design retained 93.8% of model-adjudicated true candidates but failed its pre-registered non-inferiority floor of 0.90.

arXiv cs.CR·Sep 14, 17:13 UTC · 1d agoResearch

38

Signed Rescue Routing: Harm-Aware Cascades for Efficient LLM Inference

Signed Rescue Routing improves LLM cascade efficiency by predicting when a larger model actually corrects a smaller one rather than uncertainty.

Signed Rescue Routing (SRR) is a budgeted cascade method that separately predicts rescues and regressions when escalating from a small to a large model, ranking requests by their signed difference. The authors prove this signed conditional gain is Bayes-optimal under a fixed escalation budget and add only a lightweight two-head router needing small-model output statistics at deployment. Evaluation with Qwen3-4B and Qwen3-8B on MMLU, HellaSwag, and ARC-Challenge shows better accuracy-compute tradeoffs than entropy routing and learned error predictors.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:31 UTC · 8d agoAI research

22