Top AI Papers of the Week: The Harness Era in Agents

The weekly roundup published by the DAIR.AI team on x.com sums up where AI is heading in a single sentence: the race is no longer about the bigger model but about the program around the model. The list holds ten studies, and seven of them address this exact layer, the harness, the software that organizes model calls, tool use and information flow. Most of the full texts in the roundup sit openly on arXiv, so every claim can be checked at the source. This piece opens up those ten studies one by one in plain language.

Training the program beats scaling the model

Topping the list is CMU's Harness Learning, and the idea is surprisingly simple: the solver model stays frozen while the surrounding harness program gets revised. A small proposer model reads the execution report and suggests a program edit to the harness; the proposer is trained with reinforcement learning, rewarded by the revised harness's task score. The headline result: a 4B proposer beats its 35B teacher at single-step revision. More important still, the skill transfers; a proposer learned on one question set keeps improving harnesses on sets it never saw.

From Microsoft and collaborators comes CorpusMap, the second study, pressing on a sore point of enterprise search: agents rediscover document links on every query. CorpusMap moves that work before the query; every recurring entity gets a page linking to each document that mentions it. The original documents stay put while the map is shared across queries. Across seven models and three enterprise sets, answer quality rises 6.4 to 11.7 points while input tokens fall 34 to 57 percent. That the map can be built with no model calls and updated as documents arrive makes it practical for real deployments.

Trust evidence, not consensus

Google's VeriHarness questions a common habit in long-horizon agent work: sampling a few rollouts and trusting the majority. The authors argue consensus can hide a shared error while disagreement often points at the correct alternative. The fix turns the very same generator model into a verifier equipped with evidence tools: one role tests rival claims against environment evidence, another probes the claims everyone agrees on. Topping the selection scores on five work benchmarks, the method adds more than 6 points over a single rollout on two flagship models. Nearly 26 thousand rollouts were released as a dataset, announced through the VeriHarness project site; the study's HuggingFace record widened the discussion further.

The fourth stop is HERMES, changing how we look at software engineering: every repository component gets a resident language model that knows its own job. These Dev-Primitives reason in natural language, talk to each other and edit their own parts in place. Which units to wake is decided from dependencies, and test failures are traced back to the responsible unit. On whole-repository migration the same model's score jumps from 6.5 percent to 31 percent, with an average gain of 12.4 points across four software tasks. With strong routing, 8B components stay within 4.5 points of the flagship model while cutting costs by more than a quarter.

One model, three roles: a team that trains itself

Sakana AI's MASS tries self-improvement with no outside checker: a single base model plays team member, evaluator and optimizer at once. The model proposes multi-agent work plans, runs them, picks the best by its own scoring, then fine-tunes itself on that team's traces before the cycle restarts. After two cycles, per-token performance climbs 1.2 to 1.6 times across four open-ended areas. A student raised on team traces beats a single-agent student that saw 1.4 times more traces. The study's code was shared on GitHub, and the Sakana project page walks through the method step by step.

Meta's Sharpening Tax paper prices the hidden bill of RL fine-tuning: single-shot scores rise while solution coverage shrinks. Raw models with a light surrounding program turn out to be surprisingly strong agents; fine-tuned versions win on one try but cannot reach, over many tries, tasks the raw model solves. Across 42 model pairs and three agent tasks the tax shows up in most setups and grows with size. The proposed fix, PTGS, sets sampling temperature from task difficulty; it improves both the single shot and the coverage. The DAIR academy page lists this study with an illustrated summary.

NVIDIA's VERA fixes the environment problem of long-horizon agent training: most environments score only the finale, while training needs areas that pause at any stage and score from evidence. The method takes ready execution traces and turns each into a sandbox with rules and runnable checks; over 9 thousand judged areas enter the training bank. Training alternates between model weights and harness skills. At the 9B scale it beats the strongest single-axis baseline by more than 10 points in two areas; at 27B the scores reach 71.6 and 80.7. The full text is out on arXiv.

Measuring the yield of learning

Agent Plasticity, joint work from Meta, Berkeley and Princeton, asks the right question: the model with the top score, or the one turning experience into gains most cheaply? The metric freezes weights, starts each round from a blank context, and divides held-out gains kept through earlier rounds' notes by cost. In chess, Go and Hex one model sprints to the final while another learns more per unit cost; in an infiltration game a single model jumps 66 points on a four-figure budget. Slow learners turn out never to reuse the notes they wrote. The project page is live on GitHub.

Sakana AI's Continuous Memory Machines study gives recurrent models two separate memories: a short-term one tracking current activity and a long-term one holding what will be needed later. An attention-based reader reads and writes both at every step. On copy, associative recall, sorting and maze tasks it passes predecessors such as LSTM, DNC and RMC, keeping generalization to longer inputs. Attention maps show the model using long-term memory for algorithmic work and skipping it when unneeded. The Sakana finding says squeezing short and long into one vector is no longer enough.

The closing piece, MIMESIS, looks at training's forgotten half: the user. Ready-made assistant models act too cooperatively in the user role; a fixed agent scores more easily against these fake users than against real people. MIMESIS is a 9B user simulation learned from real human conversations and 13 behavior patterns, beating the strongest flagship model by 13.4 points on behavioral fidelity. Freeze that simulation, train an agent against it, and generalization improves across all nine fresh user types. An extra step turning the simulation's private reasoning into token-level feedback grows the gain further. The full text is open on arXiv.