Search and filter through extracted claims from AI researchers.
Showing 341-360 of 587 claims in topic "agents"
AREX sustains recursive self-improvement over long horizons by learning an autonomous context-update tool that compresses growing interaction history into compact improvement state
Research agents should recursively improve answers by verifying intermediate results and using partially verified state to guide subsequent refinement, rather than simply searching longer
Deep research requires agents to find answers that jointly satisfy multiple constraints, with a discovery-verification asymmetry where verification can be decomposed into tractable constraint-wise checks
Large Language Models excel at natural language understanding and generation but remain unreliable for multi-step logical reasoning, especially in safety-critical or compliance-sensitive domains.
Neuro-symbolic approaches that couple neural models with external symbolic engines can address the gap in multi-step logical reasoning, though most integrations lack standardized interfaces.
In long-horizon LLM agent reinforcement learning, weak policies often repeat similar failures, producing uninformative rollout trajectories that limit effective policy optimization.
Existing skill-centric methods remain centered on the skills themselves rather than being designed as adaptive training-time support for the evolving policy.
OpenAI should release detailed transcripts from the Hugging Face hacking incident for the field to learn from
The Hugging Face hacking incident raises questions about whether the top-level agent knew about the hacking or if there was value drift between it and its subagents
Existing agent memory system implementations suffer from architectural fragmentation that couples different lifecycle stages, entangles evaluation logic with datasets, and provides limited support for heterogeneous memory types
MemTools enables interchangeable assembly of memory components across different systems through declarative data contracts and provides unified coordination of symbolic, neural, and multimodal memory representations
Existing image restoration agents store knowledge as static tool descriptions, manually defined degradation priors, or unstructured textual summaries, which limits knowledge accumulation and revision over long-term experience
Causal-AgentIR organizes restoration experience as causal graphs connecting degradation patterns, regions, tools, actions, quality changes, and preferences, enabling better knowledge organization than isolated textual records
Automated research systems are prone to silent failures where analysis code executes successfully yet relies on invalid causal assumptions
Explicitly encoding causal design principles and using synthetic data generation with known ground-truth effects can make silent failures in automated research systems visible
Problem formulations that prioritize rare samples with greater progress potential are more likely to guide efficient search processes
LLM-based automatic problem formulation methods mainly focus on design-intent alignment and overlook search process efficiency
Loops were a short-lived patch for models that couldn't reliably keep working on long problems until hitting a defined goal
Fable, GPT-5.6, and likely Kimi K3 can work on long problems until hitting a defined goal out of the box without assistance
Tencent WorkBuddy Bench provides a multi-domain evaluation framework for coding agents across Code, Web, Office, and Security domains
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 8,952 pending.