Search “Meta”

Meta paper finds reinforcement learning post-training improves first-attempt success rate but reduces coverage across multiple attempts

A new Meta paper finds that reinforcement learning post-training makes agents perform better on first attempts, but fewer tasks can be solved within up to 128 attempts. The study tested 14 sets of base models and post-trained models on 3 agent benchmarks, and built the PTGS sampler, which raises temperature for difficult prompts and lowers it for easy prompts. On WebShop, the gemma-4-31B base model solved over 85% of tasks within 128 attempts, while the post-trained version solved 56%; the authors' proposed Sharpening Tax was positive in 36 of 42 model-benchmark combinations, meaning coverage shrank. The authors suggest that when selecting models for retryable agents, one should measure how many tasks the model can ultimately solve, rather than looking only at first-attempt accuracy.

Princeton and Meta propose Self-Direction: storing agent error-correction experience as reusable LoRA adapters

A new paper from Princeton and Meta proposes the Self-Direction method, which lets a failed agent reflect on the current run and trains the lessons into a LoRA adapter about 0.02% the size of the model; a router then loads the closest adapter when a similar failure is encountered. On Qwen3.6-27B, the GAIA task success rate rises from 38.8% to 55.8%, outperforming in-context reflection and the Skills.md file approach. Of 185 errors that had occurred previously, 67.3% of fixes were still effective by the third encounter, versus 42.9% for simple retries.

Meta Puts JiTTests into Production: LLM-Generated Tests Catch Defects Before Code Is Merged

Meta has published the production deployment results of Catching JiTTests, a system that uses LLMs to automatically generate executable tests for pending code changes, exposing defects before the code is merged. The system employs two workflows, intent-aware and dodgy diff, and uses an LLM-as-Judge ensemble together with a rule-based method, RubFake, to filter false positives, while reducing manual line-by-line review through a human-on-the-loop approach. Meta says 589 engineers have participated, the true positive rate is 60%, and in the first half of 2026 it caught 74 severe privacy SEVs at the diff stage, reducing the overall incidence of such issues by 20%.

Meta's new paper proposes RSD: agents can learn from unlabeled historical trajectories

Meta's new paper proposes the RSD method, which lets AI agents review each step after knowing the task outcome and learn from historical attempts that no one labeled as right or wrong. On ScienceWorld, Qwen3-14B's success rate rose from 12.9% to 20.6%; on ALFWorld, Qwen3-8B reached 76.6%, surpassing all tested methods, including those that received rewards or standard answers. The authors recommend keeping agents' unlabeled run records (including failures) and regenerating the review notes each round, because when notes were frozen at round 3 the ALFWorld success rate stayed at 59.7%, while updating notes each round gave 67.9%.

Meta distills 30B model Muse Glimmer, capable of running agents on a single consumer-grade GPU

Meta distilled the dense 30B model Muse Glimmer from the larger Muse Spark, training it on synthetic agent trajectories in simulated applications; its 4-bit weights are under 20 GB, allowing it to run on a single RTX 5090. On the multi-tool benchmark MCPAtlas it scores 75.5, higher than Gemma4-31B's 54.2 and Qwen3.6-27B's 62.5; however, on OSWorld-Verified it trails Qwen3.6-27B's 75.6 with a score of 65.9.

Google paper proposes OO-Harness: evolving specialized agent harnesses by task type

A new Google paper proposes OO-Harness, which evolves agent harnesses by code class: each specialized harness inherits from a parent class and changes only the parts that its task cluster needs adjusted, after which a router selects a specialized harness or keeps the base result based on the base run's outcome. On AutomationBench business workflows, this method scores 77.41%, versus 61.62% for the base harness and 63.92% for the global method Meta-Harness; on unfamiliar applications it retains the base harness and maintains a 59.18% pass rate, while Meta-Harness's global harness drops to 44.90%. The authors recommend using the base harness as a fallback for mixed workloads and routing to a specialized harness only when a run exposes a known problem.

Meta's new paper proposes the ProPerMem benchmark: proactive assistants struggle with autonomously judging when to interrupt

A new Meta paper points out that the failure point of proactive assistants (such as AI glasses) lies in autonomously judging when to speak, not just what to say. The team built the ProPerMem benchmark, using first-person video paired with interference-rich, full-year memory averaging 618K tokens, and featuring human-verified trigger windows. All baselines scored at most 15%, even frontier models using retrieval-augmented memory; when oracle evidence was provided within the window, GPT-5.3-Chat reached 96%, but when it was left to choose the timing itself, this dropped to 38%, and to 15% after adding retrieval. Grasping both content and timing while reducing trigger frequency can raise accuracy from under 10% to 42.4%.

English translation is being prepared.

Meta releases Muse Glimmer technical report, a 30B open-weight model targeting agent tasks on consumer-grade hardware

Meta releases the Muse Glimmer technical report, a 30B open-weight model targeting agent tasks on consumer-grade hardware. At 4-bit precision, the model can run at practical speeds on a single consumer-grade GPU in a Mac or PC, and it leads open-source models of the same scale in tool use and multi-step planning. The report covers the model architecture, distillation recipe, quantization, and more.

Meta releases technical report for Muse Glimmer 30B, an agentic model for consumer-grade hardware

Meta has released the technical report for Muse Glimmer, a 30B open-weight model aimed at agentic tasks on consumer-grade hardware; it runs at 4-bit precision on a single consumer-grade GPU and leads open models of the same scale on tool calling and multi-step planning. The relayer adds that the model surpasses Qwen3.6-27B and Gemma4-31B on some metrics, is about 17GB after 4-bit quantization, targets devices with 24GB of VRAM, and uses DFlash speculative decoding to reach about 233 tokens/s on an RTX 5090, roughly 3.1 times faster than ordinary decoding.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Anthropic Report: AI Agents' Work Hours Reach 6 Times That of Humans, Share of Junior Tech Job Openings Falls to 9%

Anthropic released a report stating that by August 2026, the working hours of interactive AI agents deployed by employees had reached 6 times that of human employees, and in March, the company-wide total agent work hours exceeded human work hours for the first time. The report shows that the share of junior employees in the technology department fell from 21% in early 2025 to 14%, the share of junior tech job openings fell from 27% to 9%, and engineers using coding agents averaged 5 times as many code commits as they did 6 months earlier. The report also says its revenue per employee is 2 to 4 times that of Meta and Google during their high-growth periods, but Epoch AI cautions that the related estimates carry uncertainty.

English translation is being prepared.

English translation is being prepared.