DAIR.AI

https://x.com/i/article/2104259869402599424

The Top AI Papers of the Week (September 21 - 27): - HySparse2 - EvoOntology - Harness-Zero - JEV-as-a-Judge - Wiki Foundation Model - Self-Organizing Agent Teams - Self-Improvement via Fast Tree-search Read on for more:

Banger paper from Microsoft on prompt optimization. (bookmark it) The claim that a coding agent reading your logs beats GEPA at prompt optimization The overall finding is that you want to give a coding agent your full set of agent logs and let it write the analysis code, instead of running a search loop over small batches of trajectories. CASD has an off-the-shelf coding agent compute statistics over the whole trajectory corpus, find recurring failure modes, read representative episodes and write the findings as rules in one prompt. It needs no environment access and no validation data. Across ALFWorld, tau2-bench retail and telecom, and Spreadsheet Bench-Verified, one pass improves the unoptimized baseline by 16.6 points on average. GEPA improves it by 10.9 and SkillOpt by 5.3. Each optimized prompt costs about $1.60, more than 22x cheaper than validation-gated search. Paper: https://academy.dair.ai/papers/coding-agents-are-strong-prompt-optimizers-2609.26261

Banger paper from Salesforce AI Research on agent memory. The overall finding is that you want to store raw trajectories and decide what to extract from them when the next task arrives, instead of summarizing each run when it ends. Just-in-Time Memory uses a curator that reads the retrieved traces together with the new task and writes a short memory payload for that task. Because the payload is used right away, the curator can be trained on whether that same task succeeds. On ALFWorld, WebShop and tau2-bench it beats the strongest baseline by 16.2, 16.3 and 3.9 success-rate points. Even the untrained curator matches or beats memory that is written when a task ends. Paper: https://academy.dair.ai/papers/just-in-time-memory-learning-to-curate-task-adaptive-memory-for-llm-agents-2609.27334

Great paper from Salesforce AI Research on RL for multi-turn tool use. The finding is that you want to train the one call where the action changes the outcome, instead of spreading reward across the whole trajectory. (bookmark it) When reward depends on later turns, much of its variation comes from what happens downstream. Critical-State RL uses nested sampling to separate the reward variation caused by the current action from that noise, then trains only the selected call with contextual-bandit updates. On BFCL v4 missing-function tasks, training the selected turn adds about 14 points. Training the other candidate turn leaves accuracy flat or lower. Paper: https://academy.dair.ai/papers/critical-state-rl-diagnosing-trainable-states-for-multi-turn-tool-use-2609.24985

An interesting idea is to continually train specialized models on skills. This paper explores that idea. They propose SkillGym which turns human-written skills into 2,756 training environments with code-based checkers, then fine-tunes on 8,364 successful trajectories. After fine-tuning, Qwen3.5-35B-A3B running in Claude Code gains 19.10 points on Terminal-Bench 2.1 and 28.13 points on skill-assisted SkillsBench v1.1, where it reaches 51.47%. That is above the reported scores for Claude Sonnet 4.6 and GPT-5.4 Mini. With no skills loaded, the trained model beats the base model that has the skills in context. Paper: https://academy.dair.ai/papers/skillgym-internalizing-human-skills-into-llms-for-real-world-problem-solving-2609.27717

Interesting new paper from Google DeepMind. Studies agents that follow bad advice from users. Users often suggest a fix that sounds right and is wrong. XYEval adds one confident, misleading hint to tasks from tau2-bench, SWE-bench, Terminal-Bench, HLE and MCP-Atlas. The task and its correct solution stay the same, so any drop comes from the agent taking the advice. Scores fall by up to 46.7% relative across Gemini, Claude Opus 4.8 and GPT 5.5. The drops are larger on easier benchmarks, which suggests more capable models will not fix this on their own. Agents often disagree with the hint in their reasoning, then follow it anyway without telling the user. Compliance shows up almost only in failed runs. A system prompt that warns about the XY problem helps on single-turn tasks and leaves large drops on multi-turn ones like tau2-bench and SWE-bench Verified. Paper: https://academy.dair.ai/papers/xyeval-agents-say-yes-to-bad-advice-2609.23939

The era of omni agents is upon us. This is a great report by the Qwen Team on their omni-modal agents. They present Qwen3.8-Omni-Flash, a natively multimodal model trained for long-horizon agent tasks across text, audio, and video, such as video editing and long-form audio and video translation. It uses the sparse mixture-of-experts design of Qwen3.8-Next with a context window of one million tokens. A co-training strategy keeps text performance while carrying agent skills over to audio and video tasks. Two open-source frameworks come with it. Qwen-MM-Plugins adds audio and video support to existing agent harnesses, and Qwen-Live-Harness handles real-time multimodal interaction with context and memory management, tool use, and sub-agent delegation. Paper: https://academy.dair.ai/papers/qwen3-8-omni-towards-native-omni-modal-agents-2609.25611

https://x.com/i/article/2111861490571485185

Must-read paper integrating decision models like Jev into harnesses. They distilled an open 4B decision model, mostly from Jev's answers on the agent's own states, that beats Jev inside that agent and is much cheaper to run. Their harness, JevAgentHarness, sends each routine choice to the decision model and uses its answer only when the confidence clears a threshold fitted on the agent's own trajectories. Below the threshold, the LLM decides as usual. With the fitted threshold, tool and argument cascades cut cost by 14.2% without changing the outcome of any of 500 tasks. A fixed threshold of 0.5 lost 21.5 points. They also train AgentDecision-4B on states the agent reaches after its own mistakes, with Jev's answers as labels. On tasks unlike its training data, the agent solved 91.5% with it and 82.0% with Jev. On familiar tasks, 4B models trained this way made agents 81.1 to 89.0% cheaper and up to 2.5 times faster while staying within 3 points of success. Paper: https://academy.dair.ai/papers/jevagentharness-benchmarking-deploying-and-training-decision-models-that-share-c-2610.21412

Brilliant paper on agentic continual learning from Meta and MIT. (bookmark it) Agent memory harnesses usually only write forward. A lesson goes into the bank and is never checked against whether it helped. This work proposes ReHarness, which rewrites the reflections an agent retrieved for a task after it sees how that task turned out. Add and delete operations keep a reflection bank tidy, but they ignore downstream outcomes. ReHarness uses each finished trajectory as feedback on the specific reflections that guided it, so overgeneralized or outdated lessons get corrected. Across three continual-learning benchmarks with 16 tasks and three backbones, it beats raw-context, memory, reflection, and prompt-optimization baselines, and the gains stay significant after multiple-comparison correction. https://arxiv.org/abs/2610.12668 Chat with Paper: https://academy.dair.ai/papers/recurrent-harness-for-agentic-continual-learning-2610.12668

Insightful paper on recursive self-improvement with a frozen LLM. (bookmark it) Memento 3 clears every level of all 25 public ARC-AGI-3 games and uses 44% of the human action count. The underlying LLM is never updated. How it works: The agent writes what it believes about the environment as a natural-language rulebook, then compiles that rulebook into code that predicts the next state. A new version is kept only if replaying the code reproduces every transition the agent has observed, cell for cell. In an Atari Pong case study, the learned model drives a controller that wins 21 to 0 in three games without further LLM calls. https://arxiv.org/abs/2610.11794 Chat with Paper: https://academy.dair.ai/papers/memento-3-model-based-recursive-self-improvement-through-reflective-rulebooks-2610.11794