Blog

IF-MAP prioritizes agent improvement based on system potential Oct 24, 2026 · Multi Agent Systems · Optimization

The IF-MAP method identifies which specific agent in a multi-agent system should be upgraded or re-prompted to maximize overall system-level performance across various workflows.

RATED Cuts Faulty Tool Selection From 86.2% to 2.9% in Simulation Oct 12, 2026 · Rated · Tool Retrieval

RATED reduces faulty tool selection from 86.2% to 2.9% in a simulated 30% tool-outage benchmark. Live production reliability remains unmeasured.

RoboJEPA establishes scaling laws for multi-embodiment robotic world models Oct 8, 2026 · Robotics · Scaling Laws

RoboJEPA uses a Joint Embedding Predictive Architecture trained on 12 robotic embodiments to prove that imagination error follows a second-order power law in compute.

SearchJev separates search decisions from autoregressive language generation Oct 6, 2026 · Agentic AI · Search

SearchJev uses a calibrated System-1 model to handle search agent decisions, improving answer accuracy from 45% to 54% on BrowseComp-Plus while reducing active search time.

Agentic verification improves long-horizon task performance over single rollouts Oct 5, 2026 · Agentic AI · Verification

VeriHarness increases model performance by using a generator as an agentic verifier to resolve disagreements and challenge consensus through environmental evidence and reusable skills.

Keyword Harnesses Fail Open in Small Models Oct 3, 2026 · Language Models · Research

A keyword-based evaluation credited tool mentions without valid calls in a tested small-model pair. Strict structural checks exposed the false positive; this does not show deliberate deception by the models.

AutoCompact improves coding agent pass rates via adaptive context compaction Oct 2, 2026 · Coding Agents · Context Management

AutoCompact increases SWE-bench Verified pass rates by 9.2% by training agents to replace noisy exploration history with a clean working state instead of relying on large context windows.

AEWM: Better Agents Through Active State Revision Sep 24, 2026 · Agents · World Models

The new Agent-Editing World Model improves agent reliability by actively revising noisy reasoning and actions, preventing task-state contamination.

Memory Decision Layer: A White-Box Fix for LLM Hallucinations Sep 21, 2026 · LLM · Memory

On September 18, 2026, Yiming Zhang and colleagues introduced the Memory Decision Layer (MDL), a zero-parameter controller that tackles conflicting memories in LLMs by decoupling confidence from consistency.

Pruning Quietly Breaks Smart-Home Agents Sep 16, 2026 · Pruning · Smart Home

A September 2026 study of 19,500 smart-home tool-calling instances finds that dense models shatter under pruning while mixture-of-experts systems hold up — and the field isn't measuring degradation correctly.

ESPO beats GEPA on prompt length and accuracy by stabilizing selection Sep 4, 2026 · Espo · Prompt Optimization

On seven NLP benchmarks, ESPO improves average accuracy by 3.76 points over GEPA with prompts 47% shorter.

Coding Agents Fail 95% of Whole-Repository Refactors Aug 25, 2026 · Coding Agents · Swe Refactor Bench

The SWE Refactor Bench paper shows coding agents can't handle long-horizon tasks, with only 28 out of 520 runs passing all evaluation stages.

Execution Scores Hide Agent Command-Path Failures Aug 14, 2026 · Agents · Benchmarks

QuoteBench found that current benchmarks for LLM coding agents miss critical failures by only checking the final output, overlooking errors in command serialization and re-parsing.

DelusionEval tests how longer context affects chatbot safety Aug 7, 2026 · Delusioneval · Context Window

DelusionEval uses 589 conversation histories from 18 participants.