Adversarial training with large step sizes shows persistent implicit bias
Large constant step sizes in adversarial training preserve implicit bias gap with standard training and exhibit two-stage sharpness dynamics.
Stories per day · 30 days
2 RSS
Relatedllm-agents3adversarial-training2benchmark2function-calling2optimization2tool-use2adversarial-attacks1adversarial-robustness1agents1certification1cipr1concept-based1control1distribution-shift1interface-drift1learning-to-defer1memory-poisoning1mixture-of-experts1
31 stories, newest first · one card per story
Stories tagged #robustness
H-infinity controller improves LLM steering
H∞ controller enhances robustness for out-of-distribution LLM steering.
Robust memory manager rejects poisoned LLM agent inputs
One-shot fine-tuning on synthesized episodes with changes and attacks more than doubles evidence-conditioned accuracy against poisoning and outdated facts.
Concept-based attacks target neural-network logic, not class labels
Retraining on the attacks kept classical robustness gains while better removing semantic shortcuts.
NTBench shows higher success for near-target prompt injection
On language tasks, incorrect-target success rises 9.3 to 35.8 points versus a fixed-target baseline.
CertMoE certifies MoE LLM robustness under dynamic routing
Existing fixed-graph certificates were violated more than 30 percent of the time; routing-aware checks cut that to zero in tests.
RoboCasa-Shift measures stage-aware robustness of household VLA policies
GR00T N1.5 ran 1,040 rollouts; equal success drops hid different first-stage failure rates.
Robust Function Calling for LLM Agents Under Interface Drift
SchemaShield-Lite restores Qwen3.5-9B function-calling accuracy after tool-interface drift without retraining.
One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness
CIPR trains LLM agents to stay consistent across renamed or notationally changed interfaces.
REASON proposes two-loop residual analysis for learning
The two-loop model beats the fast loop at about half the medium loop's computation, including under unseen corruptions.
Predicting Alignment Generalization with Value Representations
Activation-based value representations predict LLM alignment generalization far better than textual descriptions of values.
Black-Box Adversarial Patch Attacks on VLAs via Ancestor VLM Exploitation
Adversarial patches optimized only on ancestor VLMs substantially cut success rates of deployed vision-language-action models.
On the Intrinsic Limited Robustness of Latent-Based Watermarking
Researchers argue latent diffusion watermarks are intrinsically fragile to rotation, scaling, and translation.
When Tools Lie: Reliability of Mathematical Agents Under Corrupted Tool Feedback
Corrupted tool feedback drops math-agent accuracy from 100% to 72.4%; mandatory reflection restores 100%.
Incidental information contaminates patient notes and disrupts clinical reasoning in large language models
Frontier LLMs insert small talk into 35% of clinical notes and sometimes treat it as clinical fact.
Generalization Is Stability, Not Accuracy: Multi-Axis Evaluation of LLMs
SAGO scores LLM generalization as stability across input variants, not single-format accuracy.
MEND: Label-Free Detection, Localisation, and Correction of Latent Hallucination in World Models
MEND detects and localizes silent latent hallucinations in frozen world models using a single score field.
Whitening Improves Robustness to Spurious Correlations in Linear Probes
Whitening linear-probe features reduces reliance on spurious correlations without labeled data or known confounders.
WISE-ATTA: When to Ask for Labels in Budgeted Active Test-Time Adaptation
WISE-ATTA schedules scarce labels during test-time adaptation and matches prior methods with fewer annotations.
It's Not What the Image Shows: Irrelevant Context Destabilises VLM Judges Without Informing Them
MIST shows irrelevant images change VLM judge labels about as often as aligned images, without improving agreement.
Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing
Self-compensating VLAs adapt online to robot execution errors, raising task success by over 30 points.
BaRe-Mem: Bayesian Reliability Memory for Robust and Adaptive Agent Consultation
BaRe-Mem tracks advisor reliability so multi-agent systems resist misleading consultations better than debate or voting.
Mind the Spike: Mechanisms and Brittleness of Visual Massive Activations in Large Vision-Language Models
Visual activation spikes in vision-language models are brittle and can be created or removed with tiny image perturbations.
Towards Understanding LLM-Based Log Anomaly Detection: An Empirical Study of Performance, Efficiency, and Robustness
Study finds LLM adaptation strategy, not scale alone, drives log-anomaly detection gaps, cost, and noise robustness.