Explicit reliability modeling cuts hallucination. Σ‑Mem stores symmetric competence states for peer agents, while LedgerMind adds a Structured Evidence Ledger that forces reasoning to cite provable tool outputs. Both pipelines let us evaluate answers systematically and keep track of where information came from [1], [2].
Why it matters: provenance lets developers audit LLM decisions and gives end‑users confidence that an answer is backed by traceable evidence rather than a model’s guess.
Reinforcement learning teaches agents when to call external tools—such as GUI or CLI planners—and how to absorb specialist perception modules into their own weights. Experiments run over >100‑turn trajectories show order‑of‑magnitude cost reductions and higher success rates on heterogeneous tasks [3].
Why it matters: adaptive tool selection turns costly, brittle pipelines into leaner, more reliable agents that can operate in open environments without manual prompting.
Hybrid pipelines combine persistent native memories embedded in model backbones with a cheap lexical pre‑retrieval stage (BM25). The InMind benchmark isolates retrieval bottlenecks; results show that BM25 dominates as corpus size grows, while the agent’s sequential search only helps on very small datasets [4].
Why it matters: faster, cheaper retrieval preserves latency budgets for interactive agents and makes large‑scale code or document assistance feasible.
LedgerMind structured evidence ledger – Constrains LLM reasoning to provable tool outputs, boosting answer accuracy and trajectory faithfulness [2].
Meta‑evolution of program‑evolution operators – Uses a meta‑evolutionary loop to discover new program‑mutation operators automatically; this enables recursive self‑improvement and sets a new state‑of‑the‑art on complex software‑engineering pipelines [5].
MAPD protocol distillation with RL – Merges structured protocol distillation with reinforcement learning, transferring proprietary reasoning skills into open models and delivering notable gains on math and planning benchmarks [6].
SpatialCLI internalizes specialist perception modules – Trains VLMs to invoke external spatial tools during inference, then distills the knowledge so the model no longer needs the tools at test time; zero‑tool performance on navigation tasks improves dramatically [7].
Residual contamination inflates macro‑F1. Dynamic fact‑checking benchmarks reveal that leftover contamination can add up to 11 points to macro‑F1, exposing a flaw in current evaluation pipelines [8].
Lexical BM25 dominates at scale. Across increasing corpus sizes, simple BM25 retrieval consistently outperforms more complex agentic search, confirming its value as a baseline for large datasets [4].
PerceptionBench ceiling at 60 % accuracy. Ten atomic visual skills are isolated; no multimodal model exceeds roughly 60 % on these tasks, highlighting a substantial gap in reliable perception [9].
CADENCE dense partial rewards improve reasoning. By adding coverage‑adaptive scheduling and dense partial rewards, CADENCE lifts performance by 3–10 percentage points across mathematics, QA, and planning benchmarks [10].