LLM agents
13 published articles
Fisher-R1 | Hypothesis testing
AI agents run flawless statistics and still draw the wrong conclusions
Agents that automate hypothesis testing can execute analyses flawlessly yet reach wrong conclusions, because most benchmarks never check whether the p-value is valid. A 425-task benchmark quantifies the gap, and an RL-trained open-weight model closes a chunk of it.
2026-08-19
AI Agents
PsychoAgent gave AI agents emotional memory. The raters found no edge
PsychoAgent gives LLM agents a separate affective memory so emotionally salient, conflict-laden traces can outweigh merely topical ones. In three conflict scenarios it retrieved more conflict-critical memories than both baselines (0.933 vs 0.500 and 0.667), yet five blinded raters found no significant quality edge.
2026-08-19
AI Research
Stop copy-pasting skills: SkillZip's evaluation-free fix for bloated agents
Self-evolving agents append fixes until the same rule appears across several branches. SkillZip compresses their skills without evaluation rollouts, by finding the shortest faithful structural explanation. One-shot and Zip-on-Write modes, and what the abstract leaves unproven.
2026-08-15
AI agents
Treating SOPs as code: why compilation alone lifts strong agents by 16 points
New research from Hong Kong and mainland China demonstrates that compiling SOPs into executable pseudo-code and running them on a stack-paged virtual machine cleanly separates capable agents from brittle ones. The work yields a precise deployment rule: compile first, page only after a model-level discipline check.
2026-08-04
LLM Agents
The regression tax: why loading LLM agents with skills can backfire
A new study shows that adding procedural skills to LLM agents does not always help, it can introduce regressions, where tasks previously solved without skills are failed after skills are added. The research identifies three causes and argues that reliability depends more on grounding and verification than on the skill itself.
2026-08-03
Artificial Intelligence
Research idea generation gets 3.89x better with IDEAgent
IDEAgent uses lineages, multi-objective feedback, and sequential memory to balance quality and diversity in LLM-generated research ideas. Tested across 32 topics in 8 CS domains, it achieves 3.89x the Yield (diverse ideas above a quality threshold) of prior methods.
2026-08-02
Zero-shot AI for industrial root cause analysis
AgentRCA: zero-shot root cause analysis that explains its own reasoning
A
2026-08-01
Research
Six words that could break the AI agent safety ceiling: disrupt, validate, broker
A novel heterogeneous agent cohort architecture separates divergent exploration, runtime safety gating, and cross-domain knowledge retrieval into specialized roles. The Disrupter generates high-entropy proposals, the Validator enforces hard tool-call checks, and the Broker imports out-of-domain analogies via contrastive novelty retrieval. Execution failures are compiled into signed constraint patches called Scars, cached for future generations. In evaluations, the cohort achieved 95% remote target discovery, zero executed breaches, and 15.1% token savings from Scars, with a 55.9% cost reduction under resource constraints via credit-based bandwidth allocation.
2026-07-30
AI Research
Harness evolution looks good until you run a fair test
Automatic harness evolution is supposed to make LLM agents better, but a new paper argues many reported gains may be from overfitting to the public test set. In experiments, simple test-time scaling methods matched or outperformed evolution, and evolved harnesses showed limited generalization to held-out tasks.
2026-07-27
Agentic RL research
RL's sparse-reward blind spot meets SEED: agents write their own lessons
SEED (Self-Evolving On-Policy Distillation) lets an LLM analyze its own past trajectories, extract reusable natural-language skills from them in hindsight, and distill those lessons back into its policy during RL. The result: denser, on-policy supervision that improves success rates by up to 22% on long-horizon tasks.
2026-07-25
Long-horizon reasoning
The one framework that lets AI agents remember what they did 47 steps ago
PRO-LONG is a minimal framework that helps LLM agents retain and retrieve information across long sequences of actions. On the ARC-AGI-3 benchmark, it improved average pass rates by 18 percentage points across frontier models while using 4.2 to 5.8 times fewer tokens than existing harnesses. With Fable 5, it hit 97.4% best@2 at a total inference cost of $1,750.
2026-07-24
AI Research
The one training trick that stops AI agents from freezing in production
Current LLM agents crumble under real-world randomness. NoisyAgent exposes them to controlled noise during training, improving both robustness and general benchmark performance. The paper suggests the field has been overfitting to pristine conditions.
2026-07-17