Agents

LLM agents and the systems around them: harnesses and orchestration, tool use and MCP, multi-agent systems, agent memory and planning

600 articles · page 6 of 13

Agents

Multi-agent adversarial review boosts LLM-based SOC alert triage accuracy

The authors studied five representative LLM-based agent approaches for tier-1 SOC alert triage and found that all missed at least 40.4% of attack-related alerts. They then designed AIDA, a multi-agent framework that requires an explicit proposed decision, independent challenge, and stronger evidentiary requirements before dismissal, achieving an F1 score of 0.958 (compared to 0.371–0.744 for baselines) and reducing the false-negative rate to 3.1%.

·3 min
Agents

Autonomous agent uses identifiability analysis to decide why a model plateau has been reached

The authors introduce LLM-IDEA, an agent that couples a large language model proposer with a deterministic identifiability engine to triage plateaus in mechanistic world model learning. The engine computes the sensitivity matrix rank and certified exhaustion, enabling a three-way verdict: capability limit, resolvable within design class, or certified exhausted. On the Alien Universe testbed, LLM-IDEA under an identifiable protocol reached discovery depth at least three on 8/8 seeds versus 1/8 without the identifiability engine.

·2 min
Agents

Diffusion language models enable faster plan generation and repair for long-horizon agents

The authors introduce Plan-and-Patch, a framework that uses a diffusion language model (dLLM) to generate structured plans and repair them by infilling affected regions while preserving surrounding steps. On Natural Plan benchmark without task-specific training, the diffusion planner achieved 53.7% repair success versus 27.0% for autoregressive planners, and after task-specific training on ALFWorld and TextCraft, diffusion reduced plan-generation latency by 39-46% while maintaining similar success.

·3 min
Agents

Language models predict research outcomes across different training environments

The authors demonstrate that language models can act as Research World Models (RWMs), predicting the outcomes of experimental interventions across multiple research environments. Using 2,653 experimental records from nine pretraining, post-training, and inference environments, they show that research knowledge acquired from real experiments improves prediction accuracy for unseen interventions within the same environment and can transfer to different environments, reducing selection regret by 78% in one case. This supports the idea that experimental data can be reused to guide future research decisions.

·2 min
Agents

Multi-agent workflows enable models to recursively improve on open-ended tasks

The authors propose Multi-Agent Self-Supervision (MASS), a method where a single model generates, evaluates, and optimizes multi-agent workflows, then fine-tunes on those trajectories. Over two cycles with a Qwen3.6-27B model, they achieve 1.2-1.6x higher performance per output token on four benchmarks, and multi-agent trajectories are more training-efficient than single-agent ones, suggesting that jointly learning orchestration and subagent execution provides effective signal for recursive self-improvement on non-verifiable tasks.

·2 min
Agents

Hippocam enables LLM agents to learn from experience without parameter updates by structuring memory as nested intents and consolidating unused history.

The authors propose Hippocam, a memory architecture that organizes an LLM agent's interaction history as a tree of nested intents, consolidating completed tasks and recursively abstracting older unused experiences. Evaluated across eight measures of memory, knowledge use, and learning, Hippocam led or matched the best compared method on seven of them, demonstrating that agents can learn and evolve capabilities through their own experiences without parameter updates.

·3 min
Agents

Event-centric memory with query-aware graph augmentation boosts long-term dialogue reasoning

Liu et al. propose QGMem, a memory framework that stores dialogue histories as event-indexed atomic memory units and consolidates related units into dynamic memory traces. When a query arrives, hybrid retrieval and query-aware reranking select a compact working memory, which is then organized as a local graph and encoded into a graph token for the LLM. Across six benchmarks, QGMem consistently outperforms flat and graph-based memory systems in retrieval, multi-hop evidence composition, conflict resolution, and ultra-long dialogue reasoning.

·3 min
Agents

Frozen LLM agents learn world models via reflective rulebooks

The authors introduce Memento 3, a system that enables a frozen large language model (LLM) to continually learn explicit world models by storing revisable hypotheses in an external natural-language rulebook. The rulebook is compiled into executable code for planning and prediction, and the agent iteratively refines both representations through a reflective loop of observation, rule revision, and verification via exact replay. On ARC-AGI-3, the single-model agent cleared all 25 public games with a mean Relative Human Action Efficiency (RHAE) of 100.0 and used only 44% of the human action count; in an Atari Pong case study, it learned a feedback controller that won 21:0 in each of three evaluated episodes without additional LLM calls.

·3 min
Agents

Training SQL agents inside their execution harness improves accuracy

Yang and colleagues address a mismatch between conventional text-to-SQL training and deployment, where agents must interact repeatedly with a database before submitting an answer. Their framework trains on verified, full interaction trajectories inside the same SQL harness used at runtime, then applies reinforcement learning based on execution correctness. On Spider 2.0-SQLite, this raised execution accuracy from 15.5% to 45.2% for Qwen3-8B and from 22.2% to 54.8% for Qwen3-14B.

·2 min
Agents

Multi-agent LLM system solves hundreds of open math problems without human steering

The authors built Bolzano, a multi-agent system where parallel prover agents propose informal proofs and a verifier agent checks them, all maintaining a human-readable research state. After initial expert-guided case studies produced eight verified results, they ran the system without problem-specific guidance on approximately 3,800 open problems extracted from four sets of papers, solving about 200 of them. Among these, four solutions to questions raised in papers accepted to STOC 2026 were confirmed by the original authors.

·3 min
Agents

Step-by-step process delivery boosts LLM predictability and auditability

Schabert and Peters propose delivering Standard Operating Procedure steps one at a time via the Model Context Protocol, rather than embedding the full SOP in the system prompt. Across 15,475 trials on SOP-Bench, this step-level approach raises process adherence from 76–95% to 95–99% and reduces ungrounded correct answers from 2.1–4.5% to 0.2–0.3%, while lightweight executors gain 6.5 percentage points in grounded accuracy.

·3 min
Agents

Spoken dialogue memory jointly tracks who said what and when

The authors present PERSIST, a persistent memory system for multi-session, multi-speaker spoken dialogue that retrieves past conversational events by jointly scoring semantic content, acoustic speaker identity, and temporal state. On their new SpokenTrace benchmark, it reaches 85.08% end-to-end task accuracy and improves all-support EM@3 from 49.01% with a strong semantic retriever to 82.10%. By reusing intermediate dialogue-backbone representations, PERSIST avoids re-encoding query audio and reduces retrieval latency from 578.42 ms to 7.03 ms.

·2 min
Agents

Personal LLM agents improve recommendations using cross-platform history

The authors formalize a new recommendation paradigm in which a personal LLM agent uses cross-platform user history to mediate a platform's candidate ranking before delivering the final top-K slate. They introduce MediateRec, a benchmark with proxy and real cross-platform environments, and propose Personal Attribution Mediation Optimization (PAMO), a training method that counterfactually masks cross-platform history to allocate advantage mass while preserving platform-relative value. Experiments show that PAMO consistently improves rescue-harm balance over outcome-only reinforcement learning, yet even strong proprietary LLMs introduce non-negligible harmful overrides.

·3 min
Agents

Bidirectional protocol connects LLM agent workflows with inference engines

The authors propose HEAR, a bidirectional protocol that defines four semantic categories for information exchange between an agent harness (which manages workflow dependencies and context) and an inference engine (which handles request batching and KV-cache). Implementing two coordination policies—cache-aware runtime scheduling and role-specific inference configuration—the authors demonstrate on four benchmarks that combining harness intent with engine state improves throughput and latency without degrading task quality.

·3 min
Agents

LLM aggregators partly count restated evidence as new readings, measure shows

The authors develop a metric that converts an LLM aggregator's reported probability into the number of independent readings it implies, so a restated copy receives a weight between 0 (counts only sources) and 1 (counts every statement). Testing four LLMs across three testbeds—controlled logs with an exact Bayesian oracle, web documents with appended copies, and agent-written logs under four communication protocols—they find that a forwarded copy counts for 0.06 to 0.42 of a new reading. A one-paragraph declaration of what a copy contributes reduces copy weight to 0.08 or less on controlled logs, and a rule that has agents refer to readings instead of restating them cuts belief-implied early commitment from 11.2% to 1.1%.

·3 min
Agents

Automated observer catches fail-plausible LLM errors but misses novel ones

The authors built an automated user-viewpoint observer targeting fail-plausible silent failures in a production LLM agent runtime. The two-layer pipeline—five deterministic signals distilled from past incidents escalate to an LLM judge that must cite verbatim evidence—achieves 6/6 detection on regression incidents with 0 false positives, but 0/4 recall on novel patterns. A pre-registered live deployment over 12 days fired zero verdicts, reported as undefined precision per the protocol.

·2 min
Agents

MiniCorp: a simulator for studying AI-run companies and generating enterprise data

The authors present MiniCorp, a simulator that couples an internal world of persistent agent roles with an external world of customers, competitors, and market mechanisms. Using an e-commerce company as a demonstration, they show that agents can coordinate, adapt to market feedback, and sustain long-term strategies. The simulator generates continuous records of decisions and outcomes, and checkpointing enables replay under alternative decisions, providing counterfactual data unavailable from historical archives.

·3 min
Agents

Selective KV-cache reuse depends on request history in rolling agent workloads

The authors show that when a KV cache persists across multiple LLM calls in a rolling agent workload, the same prompt can produce different answers depending on the sequence of earlier requests. With a fixed 5% recomputation budget, document-aligned recomputation reduces answer variation across request orders from 69.0% (with CacheBlend's token top-k policy) to 26.1%. It also improves fidelity to full prefill by 34.5–52.5 percentage points while both policies achieve approximately 5.7× median time-to-first-token speedup.

·3 min
Agents

Agent improves by distilling its own verified experiences during deployment

The authors study Online Agentic Test-Time Training (OaTTT), where a deployed LLM agent executes each task once and uses the single verified trajectory as the only learning signal. They identify that directly imitating or reinforcing these trajectories destabilizes the policy. They introduce ASCENT, which self-distills verified experience by using a frozen copy of the LLM as a teacher, conditioned on the successful trajectory as privileged information, to produce stable next-token distributions that are distilled into persistent LoRA weight updates. Across ALFWorld, WebShop, and AppWorld, ASCENT improves task success and interaction efficiency as experience accumulates, outperforming online adaptation methods and transferring to held-out scenes.

·3 min
Agents

LLM tool-use mitigations can worsen hallucination in new configurations; EscapeGuard prevents this

Qi et al. evaluate five existing mitigation methods for tool hallucination in LLM agents across multiple tool configurations and discover hallucination escape: these methods reduce errors in their tuned setting but increase them elsewhere, cancelling gains. They trace this to intrinsic tool-use tendencies that models hold, which existing methods reinforce. The authors propose EscapeGuard, a training-free method that detects conflicts between model tendencies and the current configuration and redirects attention to the runtime tool set, reducing tool-selection hallucination by 9.0 percentage points and lowering cross-configuration mean hallucination by 23.7 pp.

·3 min
Agents

Reinforcement fine-tuning improves LLM social reading more than game wins

Zhang et al. let reinforcement fine-tuned (RFT) large language model (LLM) agents play social deduction games requiring hidden-state inference, social reading, and vote steering. They find that optimizing terminal win–loss outcomes does not reliably produce general social deduction ability, but RFT significantly improves social reading—inferring hidden roles from discussion and predicting others' decisions. Gains in social influence (steering votes, team approvals) require more behaviorally specific rewards.

·3 min
Agents

Shared agent memory can leak information across enterprise users

Mudgal et al. study enterprise AI agents that store multiple users’ long-term memories in a shared vector database and rely on similarity search to retrieve relevant records. Across six controlled experiments, they show that ordinary queries can retrieve semantically related memories owned by another user, even without prompt injection or system compromise. Filtering retrieved records by ownership was the only tested mitigation that consistently returned response contamination to the clean baseline.

·3 min
Agents

Preference-grouped reinforcement learning improves social adaptation in simulated conversations

Wang and colleagues introduce preference-batched group relative policy optimization, or PB-GRPO, for training conversational agents to infer and accommodate users’ unstated preferences over multiple turns. In a persona-driven simulation, the method outperformed vanilla GRPO and other reinforcement-learning baselines on an LLM-judged measure of user satisfaction, although the supplied paper excerpt does not report effect sizes.

·3 min
Agents

System that builds agent teams per query outperforms all baselines

The authors introduce SHIFT, a framework that constructs a multi-agent harness—the combination of agent roles, instructions, and tools—for each query without executing candidate designs during search. Across 9,193 tasks on six benchmarks, SHIFT-search achieves 79.9% mean accuracy, outperforming 17 baselines by 7.2 percentage points, while a cheaper variant, SHIFT-value, exceeds all baselines with 32% fewer execution tokens.

·3 min
Agents

Rewriting harness-assisted solutions improves general terminal-task performance

Li and colleagues introduce Recursive Self-Rewrite, a pipeline for turning solutions found with specialized agent controllers into supervised training data compatible with a general-purpose controller. Using Qwen-3.8-27B throughout discovery, rewriting and execution, they expand 2,001 successful source trajectories into 11,094 verified trajectories and report substantial gains over both the base model and direct trajectory finetuning.

·2 min
Agents

Auditable world-model harness achieves perfect ARC-AGI-3 score

The authors present Kepler, an open-source harness for ARC-AGI-3 that represents agent hypotheses as executable world models and validates them through retrodiction and prediction checks. Under a frozen Claude Opus 5 configuration, Kepler achieved a server-verified 100.00 RHAE on all 25 public games, with 181 of 183 completed levels using no more actions than the median-human baseline. The paper also reports three evaluation failures—source-code leakage, autonomous tool reconstruction, and silent planner repair—that score-only evaluation would have missed.

·3 min