Daniel Lenton (@DanielLenton1)

Pinned post

OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience.

On Continual-ARC, our new benchmark at @letsunifyai, none of them were able to memorize anything of value and none were able to generalize to new tasks. Training a small 9B model's weights did both.

6:52

Daniel Lenton @DanielLenton1 Nov 2

Really good read. Hadn't realized some states limit what families can spend school choice money on because checking every purchase by hand doesn't scale. Pathfinder's card checks each one against the rules instead

Amaan Verified account @amaankahmad Nov 2

x.com/i/article/2117…
View quoted post on X

Daniel Lenton @DanielLenton1 Oct 25

Pretty sure @gmi_cloud's GLM-5.3 Flash on @OpenRouter is broken, and I can't find anyone who's reported it. Since Wednesday night a fair few of its replies look like this one, which spends 32,000 tokens convinced there's an 8 in a row of 4s.

They all come back as normal 200s, so OpenRouter still has GMI at 100% uptime and nothing flags them. One of our runs went from solving 94% of its puzzles to 53% before we caught it, so that's a few days of experiments to rerun 🙄

Daniel Lenton @DanielLenton1 Oct 25

Found this today. Pretty cool that OpenResearch's agents now read their own training curves before deciding what to try next.

The search runs in rounds. The agent tries a few options for one choice, say the learning rate, and the winner becomes the parent for the next round. Each option only runs once for now, so a lucky seed can win a round (they've said as much themselves). Then the whole next round gets compared against that lucky number, so a real improvement smaller than the luck looks like a regression. And the agent's told to stop after about three regressions in a row, so one lucky winner could end a search that was still finding things.

Rerunning the winner and its parent on a few seeds before promoting it is on their list, which should fix this too. Excited to see what it finds once that's in!

alphaXiv Verified account @askalphaxiv Oct 23

OpenResearch now has @TrackioApp built in 📈

Your agent logs its experiments to Trackio, charts open right next to each run, and it reads the curves back to decide what to try next. Open source and local-first.

Running on HF Jobs, Modal, k8s or Slurm? Save a @huggingface Space and every run lands there, private Spaces included.

Check it out: github.com/alphaXiv/OpenR…

View quoted post on X

Daniel Lenton @DanielLenton1 Oct 23

Nice paper, teaching agents to edit their own context mid-task with a few special tokens (NCE), rather than leaving it to the harness. On LongCat the tool-call version, trained on exactly the same edits, came in a point below plain compaction (47.3 vs 48.3), with NCE at 53.9. Though the tool version also leaves every call and its result sitting in the history, while the native edits vanish once they run. Wonder how much of the gap is just that

Weiyang Liu Verified account @Besteuler Oct 22

Can LLMs manage their own context natively without any external harness? We are teaching LLMs the ability of Native Context Editing. Our method achieves considerable performance gain on popular long-horizon agentic benchmarks (Terminal-Bench, SWE-bench) while reducing token usage significantly.

🤔 This can be viewed as a form of verbalized context management, which leverages LLMs' inherent capabilities to optimize context. Our prior work, Verbalized Machine Learning (arxiv.org/abs/2406.04344), similarly explores context optimization to enhance downstream task performance.

Project: dengyl20.github.io/native-context…
Paper: arxiv.org/pdf/2610.25550
Code: github.com/dengyl20/nativ…

View quoted post on X

Daniel Lenton

Latest posts 5