268

Papers

10/10

Roadblocks Active

1

Connections

The JailAgent framework attacks AI agents by manipulating their internal reasoning chains and memory retrieval systems rather than injecting malicious text into user prompts — a fundamentally different threat model than most defenses assume. By extracting trigger conditions, hijacking the agent's reasoning trajectory, and tightening constraints to force compliance, attackers can bypass security measures across different models and scenarios. This matters because it shows that agent safety cannot be solved at the prompt level alone; the reasoning process itself is an attack surface that needs to be hardened.

██████████ 1.0 agent-tool-use

ReadSaveConnections

Robotic AI models (Vision-Language-Action models) that appear capable can catastrophically fail when given instructions phrased differently than training examples — task success rates collapsed from 93.33% to 5.85% under adversarially varied wording. Standard red-teaming tools suffer from mode collapse, repeatedly finding the same few failure cases and missing the broader vulnerability landscape. The DAERT framework forces diversity in adversarial testing, revealing that current VLA models are far more brittle than standard evaluations suggest.

██████████ 1.0 embodied-ai

ReadSaveConnections

Standard agent evaluation that only looks at final outcomes — not the full reasoning trajectory — misses 44% of safety violations and 13% of robustness failures, meaning current benchmarks are systematically overestimating agent reliability. Claw-Eval introduces three independent evidence channels (execution traces, audit logs, environment snapshots) across 300 human-verified tasks to catch failures that outcome-only metrics hide. The finding that video-based tasks perform significantly worse than document or image tasks also flags a concrete modality gap in multimodal agent deployment.

██████████ 0.9 agent-tool-use

ReadSaveConnections

The best available AI model scores only 29.23% on EpiBench's hard tasks, which require agents to proactively search for information, integrate multiple pieces of evidence across turns, and sustain relevant context over time — capabilities that existing benchmarks largely ignore. The process-level evaluation framework reveals not just whether agents get the right answer, but where in the multi-turn workflow they break down. This matters because real research and analysis tasks require exactly these sustained, multi-evidence workflows that current agents cannot reliably execute.

██████████ 0.9 multimodal-understanding

ReadSaveConnections

The SafeGate system adds a neurosymbolic safety check that intercepts LLM-generated robot commands before execution, rejecting unsafe instructions while accepting benign ones with high accuracy. It pairs this with Task Safety Contracts — structured rules with invariants, guards, and abort conditions — that prevent unsafe state transitions even if a bad command slips through. This layered approach addresses the practical problem that LLMs controlling physical robots can translate ambiguous natural language into dangerous actions, especially under edge-case instructions.

██████████ 0.9 agent-tool-use

ReadSaveConnections

Forcing language models to reflect on their answers in a structured format — a popular technique for improving self-correction — backfires: models achieve near-perfect adherence to the required format while completely failing to fix underlying reasoning errors. The paper identifies a new failure mode called 'structure snowballing' where the model becomes trapped in satisfying formatting requirements, crowding out actual error detection. This is practically important because constrained decoding is widely used in production systems under the assumption that structure implies correctness.

██████████ 0.9 reasoning-reliability

ReadSaveConnections

Diffusion-based multimodal language models determine their final answer at very early generation steps, before they have adequately processed visual information — essentially guessing before looking. Two targeted interventions (penalizing early answer tokens and amplifying visual grounding signals) delay this premature commitment and improve reasoning quality. This reveals a structural difference between diffusion and autoregressive models that has direct implications for how visual AI systems should be trained and evaluated.

██████████ 0.9 reasoning-reliability

ReadSaveConnections

Human financial experts substantially outperform the best AI systems on complex financial modeling tasks that average over 18 hours of skilled human labor each, with AI particularly struggling to produce client-ready outputs. The benchmark tests long-horizon computer use — navigating real software, synthesizing data across sources, and maintaining coherence over many steps — exposing the gap between AI capability on short tasks and the sustained, judgment-intensive work professionals actually do. This provides a concrete, economically grounded measure of where AI agents currently fall short in high-stakes domains.

██████████ 0.8 reasoning-reliability

ReadSaveConnections

VL-MDR replaces the opaque single-number reward signal used to train vision-language models with a decomposed set of interpretable dimensions, each weighted dynamically based on what actually matters for a given input. A visual-aware gating mechanism identifies which evaluation dimensions are relevant and adapts their weights per example, outperforming existing open-source reward models on standard benchmarks. Interpretable reward modeling matters because understanding why a model is rewarded or penalized is essential for diagnosing misalignment and building trust in AI evaluation pipelines.

██████████ 0.8 hallucination-grounding

ReadSaveConnections

Robotic AI systems that generate language descriptions of planned actions before executing them often misalign the two — the words say one thing while the robot does another — because training never explicitly connects these modalities. This paper uses contrastive learning to measure language-action alignment and then applies offline preference learning to penalize mismatched plans, tightening the connection between verbal reasoning and physical execution. Better language-action grounding is foundational for robots that need to explain or be audited on their behavior.

██████████ 0.8 embodied-ai

ReadSaveConnections