Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #8498

LLM-based agents

Related event timeline, sources and context from the news index.

EVENT TIMELINE

7

RESEARCH · 1 SOURCE · arXiv cs.AI

GRAFT: execution-guided genetic recombination for automated agent design (arXiv:2611.23509v1)

The paper introduces GRAFT (Execution Guided Genetic Recombination for Agent Design), an evolutionary search method that uses per-parent execution evidence to guide semantic crossover and integrated mutation when designing LLM-based agent architectures. Evaluated across six reasoning and programming tasks with open-weight optimizer/executor models, GRAFT achieves a mean test accuracy of 82.0%, outperforming the strongest baseline by 9.8 percentage points without requiring extra task executions.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Phronix: capability-centric learning paradigm for LLM-based agents

Phronix is a capability-centric learning paradigm for large-language-model-based agents that distills agent executions into a reusable capability store using a predictor–validator–corrector loop. The arXiv v1 paper reports that learned capabilities let agents solve recurring tasks on first attempt 93–95% of the time, produce 7.6–30.8 durable capabilities per task, reduce warm-start token costs to 0.38–0.40M tokens per task, and improve correctness and solve rates on a production I/O analysis workflow (1000 Genome Pegasus) versus a baseline harness.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

SPEM: structure-aware procedural memory for efficient skill use in LLM agents (arXiv preprint)

The paper introduces SPEM, a lightweight, structure-aware procedural memory framework that stores Atomic Skill Memory for individual skills and Skill-DAG Memory for reusable skill compositions; it uses Skill-GNN for state- and structure-aware selection and writes successful executions back as new Skill-DAGs. On ALFWorld and SkillsBench, SPEM achieves an 87.1% ALFWorld success rate and 38.38% SkillsBench reward while reducing token and runtime costs compared to strong baselines.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Testing LLM-based agent planning against formal PDDL semantics

This arXiv paper evaluates claims that LLM-based agents can perform planning by grounding natural-language tasks in hidden PDDL models that formally determine action applicability and goal achievement. Across five domains, four models, six prompting strategies and three agent architectures, the best configuration solved 78% of solvable tasks offline but only 40% online; unsolvable tasks were recognised in under 10% of episodes, and heavier prompting used 5–22× more tokens without clear benefit.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

TopoPlanner: topology-consistent task planning for LLM-based agents

The paper (arXiv:2610.07004v1) introduces TopoPlanner, a planning framework that lifts tool dependency graphs into cellular workflow complexes and uses cosheaf-consistent cellular retrieval plus multidimensional structural reasoning as topology-aware context for LLM tool-sequence generation. Experiments on four tool-planning benchmarks report consistent improvements over prompt-based and graph-enhanced baselines across different local LLM backbones for workflows with loops, merges, and reusable intermediate states.

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Before Agents Decide: Epistemic Action in LLM-Based Systems

New arXiv preprint (arXiv:2610.00511v1) brings the cognitive-science concept of epistemic actions to LLM-based agents, distinguishing three modes—acquiring missing evidence, transforming available evidence, and probing systems to elicit revealing responses—and introduces the term epistemic scaffolding for interfaces, tools, and environments that enable and make such actions auditable. The paper argues agent design should explicitly address how decision-ready evidence is produced before a decision step.

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

PFArena: benchmark of PLMs, LLMs, and agents for protein modification (arXiv:2609.28921v1)

PFArena is a new benchmark (arXiv:2609.28921v1) that provides four controlled task interfaces for protein modification, covering single-mutant generation and multi-mutant ranking under varying levels of mutation fitness data. The authors evaluate six PLMs, six LLMs, and five LLM-based agents using complementary metrics, find that PLMs excel at open-ended single-mutant generation while LLMs and agents perform better at multi-mutant ranking when target-specific fitness data exist, and report that all model families struggle as search-space size and mutation depth increase; the code and benchmark suite are released to support reproducible work.

7.0