Search “Meta”

Meta paper finds that RL post-training reduces task coverage under multiple attempts

Meta published a paper studying the impact of reinforcement learning post-training on agents' performance across multiple attempts. The study tested 14 pairs of base models and post-trained models on 3 agent benchmarks, and built the PTGS sampler, which raises the temperature for difficult prompts and lowers it for easy prompts. The results show that post-training makes agents perform better on their first attempt, but they solve fewer tasks that are solvable within up to 128 attempts. On WebShop, the gemma-4-31B base model solved over 85% of tasks within 128 attempts, while the post-trained version solved 56%. The Sharpening Tax proposed by the authors is positive in 36 of 42 model-benchmark combinations, meaning coverage shrinks. 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.

Meta paper finds RL post-training reduces task coverage across multiple attempts

Meta published a paper studying the impact of reinforcement learning post-training on agents' performance across multiple attempts. The study tested 14 pairs of base models and post-trained models on 3 agent benchmarks, and built the PTGS sampler, which raises the temperature for difficult prompts and lowers it for easy prompts. The results show that post-training makes agents perform better on their first attempt, but they solve fewer tasks that are solvable within up to 128 attempts. On WebShop, the gemma-4-31B base model solved over 85% of tasks within 128 attempts, while the post-trained version solved 56%. The Sharpening Tax proposed by the authors is positive in 36 of 42 model-benchmark combinations, meaning coverage shrinks. 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.

View story:Meta paper finds that RL post-training reduces task coverage under multiple attempts

Rohan Paul introduces the paper by Princeton and Meta on continuous improvement of AI agents

On October 30, 2026, Rohan Paul published an article introducing a new paper from Princeton and Meta on the continuous improvement of AI agents. The paper proposes the Self-Direction method: it has failed agents reflect on the current run, trains the lessons into LoRA adapters about 0.02% the size of the model, and then a router loads the closest adapter when a similar failure is encountered. On Qwen3.6-27B, the GAIA task success rate rose from 38.8% to 55.8%, outperforming in-context reflection and the Skills.md file approach; among 185 errors that had occurred previously, 67.3% of fixes were still effective by the third encounter, compared with 42.9% for simple retries. There have been no subsequent developments to this introduction so far.

Rohan Paul introduces the Princeton and Meta paper on continuous improvement of AI agents

On October 30, 2026, Rohan Paul published an article introducing a new paper from Princeton and Meta on the continuous improvement of AI agents. The paper proposes the Self-Direction method: it has failed agents reflect on the current run, trains the lessons into LoRA adapters about 0.02% the size of the model, and then a router loads the closest adapter when a similar failure is encountered. On Qwen3.6-27B, the GAIA task success rate rose from 38.8% to 55.8%, outperforming in-context reflection and the Skills.md file approach; among 185 errors that had occurred previously, 67.3% of fixes were still effective by the third encounter, compared with 42.9% for simple retries. There have been no subsequent developments to this introduction so far.

View story:Rohan Paul introduces the paper by Princeton and Meta on continuous improvement of AI agents

Meta deploys Catching JiTTests to the production environment

Meta has published the production deployment results of Catching JiTTests: the system uses large language models to automatically generate executable tests for code changes awaiting merge, 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 the rule-based method RubFake to filter false positives, while reducing manual item-by-item review through a human-on-the-loop approach. Meta says 589 engineers have participated, with a true positive rate of 60%; 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 deploys Catching JiTTests to production

Meta has published the production deployment results of Catching JiTTests: the system uses large language models to automatically generate executable tests for code changes awaiting merge, 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 the rule-based method RubFake to filter false positives, while reducing manual item-by-item review through a human-on-the-loop approach. Meta says 589 engineers have participated, with a true positive rate of 60%; 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%.

View story:Meta deploys Catching JiTTests to the production environment

Meta paper proposes RSD method to let agents learn from unlabeled trajectories

Meta published a paper proposing the RSD method, which lets AI agents, after knowing the task outcome, review each step and learn from historical attempts that no one labeled as right or wrong. Experimental results reported in the paper show: 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 also recommend retaining agents' unlabeled run records (including failures) and regenerating the review notes each round, because when notes were frozen at round 3, ALFWorld success rate stalled at 59.7%, whereas with notes updated each round it was 67.9%. Currently this work is released as a paper, with no subsequent progress yet.

Meta paper proposes RSD method to enable agents to learn from unlabeled historical trajectories

Meta published a paper proposing the RSD method, which lets AI agents, after knowing the task outcome, review each step and learn from historical attempts that no one labeled as right or wrong. Experimental results reported in the paper show: 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 also recommend retaining agents' unlabeled run records (including failures) and regenerating the review notes each round, because when notes were frozen at round 3, ALFWorld success rate stalled at 59.7%, whereas with notes updated each round it was 67.9%. Currently this work is released as a paper, with no subsequent progress yet.

View story:Meta paper proposes RSD method to let agents learn from unlabeled trajectories

Meta distills a 30B on-device agent model, Muse Glimmer

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. According to a report dated October 22, 2026, on the multi-tool benchmark MCPAtlas, Muse Glimmer scored 75.5, higher than Gemma4-31B's 54.2 and Qwen3.6-27B's 62.5; however, on OSWorld-Verified it lagged behind Qwen3.6-27B's 75.6 with a score of 65.9. There are currently no subsequent developments.

Meta paper argues that proactive assistants fail at autonomously judging when to interrupt

On October 21, 2026, Meta published a paper pointing 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 memories averaging 618K tokens, and featuring manually verified trigger windows. The paper reports that 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 after adding retrieval it fell to 15%. Grasping both content and timing while reducing trigger frequency can raise accuracy from under 10% to 42.4%. The work is currently released as a paper, with no subsequent progress yet.

Meta Releases Muse Glimmer Technical Report

Meta has released the Muse Glimmer technical report. The report introduces a 30B open-weight model aimed at agentic tasks on consumer-grade hardware. According to the report, at 4-bit precision the model can run at practical speeds on a single consumer 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 aspects including model architecture, distillation recipe, and quantization. There are no further developments at present.

Meta releases Muse Glimmer technical report

Meta has released the Muse Glimmer technical report. The report introduces a 30B open-weight model aimed at agentic tasks on consumer-grade hardware. According to the report, at 4-bit precision the model can run at practical speeds on a single consumer 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 aspects including model architecture, distillation recipe, and quantization. There are no further developments at present.

View story:Meta Releases Muse Glimmer Technical Report

Meta compresses its Agent model Muse Glimmer 30B to run on-device

On October 20, 2026, Meta released the technical report for Muse Glimmer, a 30B open-weight model aimed at agent tasks on consumer-grade hardware, running at 4-bit precision on a single consumer GPU. The reporter adds that the model leads open-source models of the same scale in tool calling and multi-step planning, surpassing Qwen3.6-27B and Gemma4-31B on some metrics; after 4-bit quantization it is about 17GB, targeting 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. There are currently no subsequent developments.

Meta distills a 30B on-device agent model, Muse Glimmer

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. According to a report dated October 22, 2026, on the multi-tool benchmark MCPAtlas, Muse Glimmer scored 75.5, higher than Gemma4-31B's 54.2 and Qwen3.6-27B's 62.5; however, on OSWorld-Verified it lagged behind Qwen3.6-27B's 75.6 with a score of 65.9. There are currently no subsequent developments.

View story:Meta distills a 30B on-device agent model, Muse Glimmer

Meta paper proposes that proactive assistants fail at autonomously judging when to interrupt

On October 21, 2026, Meta published a paper pointing 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 memories averaging 618K tokens, and featuring manually verified trigger windows. The paper reports that 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 after adding retrieval it fell to 15%. Grasping both content and timing while reducing trigger frequency can raise accuracy from under 10% to 42.4%. The work is currently released as a paper, with no subsequent progress yet.

View story:Meta paper argues that proactive assistants fail at autonomously judging when to interrupt

Meta compresses its agent model Muse Glimmer 30B to run on-device

On October 20, 2026, Meta released the technical report for Muse Glimmer, a 30B open-weight model aimed at agent tasks on consumer-grade hardware, running at 4-bit precision on a single consumer GPU. The reporter adds that the model leads open-source models of the same scale in tool calling and multi-step planning, surpassing Qwen3.6-27B and Gemma4-31B on some metrics; after 4-bit quantization it is about 17GB, targeting 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. There are currently no subsequent developments.

View story:Meta compresses its Agent model Muse Glimmer 30B to run on-device

English translation is being prepared.