264

Papers

10/10

Roadblocks Active

0

Connections

LLM-based coding agents currently treat every task as maximally complex, auditing entire codebases even for trivial fixes. The E3 framework (Estimate, Execute, Expand) first estimates task scope before committing resources, then widens its search only when the minimal attempt fails — achieving 100% success while cutting cost by 85%, token use by 91%, and files inspected by 92%. This matters because it demonstrates that smarter task-scoping, not more compute or larger models, is a primary lever for practical AI agent efficiency.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

Robotic vision-language-action models typically work from flat 2D images, losing the depth and spatial structure needed for precise manipulation. VistaVLA constructs 3D Gaussian maps that encode both geometry and semantics from multi-view inputs, then applies a Merge-then-Query summarization that compresses the token representation by 99% while preserving the spatial layout the action planner needs. The approach improves spatial reasoning in manipulation tasks beyond what 2D-only systems can achieve, at a fraction of the compute cost.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

AI sycophancy — models agreeing with users rather than remaining truthful — is not a single behavior but a family of compliance patterns whose manifestation depends on whether the interaction is single-turn, multi-turn, multi-agent, or tool-using. This survey of 126 papers finds that nearly all evaluation and mitigation work targets single-turn settings, while deployed agents increasingly operate in multi-turn and agentic contexts where sycophancy looks and behaves differently. The gap between where the research is focused and where models are actually deployed is a significant alignment blind spot.

██████████ 0.9 alignment-safety Peer-reviewed

Read

Current ML systems optimize for prediction — mapping inputs to outputs — without encoding the causal mechanisms that generate the data. This paper argues that genuine scientific discovery requires mechanistic world models: representations that capture why phenomena occur, not just what co-occurs with what. The implication is that interpretability should be a foundational design constraint, not a post-hoc analysis layer, for AI systems aimed at advancing science.

██████████ 0.9 interpretability Preprint

ReadSaveConnections

Evaluating AI memory through downstream question answering masks a deeper problem: models can score well while relying on stale, inconsistent, or unsafe internal states because right answers and correct memory management are not the same thing. MemOps treats memory as a lifecycle of explicit operations — remembering, forgetting, updating, and reflecting — and tests whether agents execute each correctly in long task-oriented conversations. All major memory architectures (long-context, retrieval-based, parametric, managed) fail to reliably handle this operational level.

██████████ 0.9 long-context Preprint

ReadSaveConnections

LLM-powered agents are being composed into complex distributed workflows without the engineering rigor that made web services reliable — composition, interoperability, governance, and trust are all being solved ad hoc. This paper maps that gap against the lessons of Service-Oriented Computing and proposes six foundational principles: harness-ability, composability, lifecycle engineering, trustworthiness by design, goal-driven orchestration, and observability. It frames agentic AI as an engineering discipline that needs standards, not just demonstrations.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

A 3B-parameter vision-language model for physical-world agent tasks now matches the performance of previous-generation 32B-parameter models, achieving top results on 19 of 38 embodied AI benchmarks. The key design choice is an action-centric capability taxonomy that organizes training data around what agents need to do — state understanding, action-transition reasoning, and sequential planning — rather than generic perception. This suggests that structured data pipeline design can close much of the gap between large and small models without scaling compute.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

Multimodal large language models can perceive UAV scenes reasonably well but systematically fail at modeling their own position and motion within those scenes — a capability called self-awareness that is distinct from, and currently weaker than, spatial cognition. Performance degrades progressively as tasks move from perception to memory to reasoning about the agent's own dynamic state. Since UAV autonomy requires exactly this self-modeling, the benchmark exposes a concrete gap between MLLM capability and deployment requirements.

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

ReadSaveConnections

Visual reasoning trained exclusively on visual demonstrations — no image-text pairs — achieves a 25% improvement on the VR-X unified visual reasoning benchmark. The VR-GRPO reinforcement learning method combines global consistency rewards with step-level logical coherence rewards, enforcing sound reasoning without task-specific tuning. The result challenges the assumption that language supervision is a prerequisite for strong visual reasoning and shows cross-task transfer from pure visual training to multimodal understanding benchmarks.

██████████ 0.8 reasoning-reliability Preprint

ReadSaveConnections

Static knowledge graphs used in retrieval-augmented generation go stale and cannot correct their own errors during inference. EvoGraph-R1 frames retrieval as a Markov Decision Process, letting an agent dynamically expand, refine, and prune a multimodal knowledge hypergraph through closed-loop actions (retrieve, expand, refine, terminate). This self-evolving approach outperforms fixed GraphRAG pipelines on accuracy, coverage, and traceability, directly targeting hallucination caused by outdated or incomplete knowledge graphs.

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

ReadSaveConnections