|
Download writeup/ABSTRACT.md from FatUnc6969/lexi-borderland-data: direct link, hf CLI and curl.
- Browser
- Download file 2.68 kB
-
https://huggingface.co/datasets/FatUnc6969/lexi-borderland-data/resolve/main/writeup/ABSTRACT.md
- Command line
-
hf download hf://datasets/FatUnc6969/lexi-borderland-data/writeup/ABSTRACT.md
-
curl -L -o ABSTRACT.md https://huggingface.co/datasets/FatUnc6969/lexi-borderland-data/resolve/main/writeup/ABSTRACT.md
2.68 kB
| # Teach the working, not the transcript: a 3B model that survives its own memory wipe | |
| Long-running agents lose their context: it is compacted, crashes, or is handed off, and the agent must continue from what it wrote down. Raw agent traces do not by themselves verify what survives this moment. We built Borderland, a deterministic 7-room environment with a memory wipe after room 3 (only an 8-line notebook survives), trap tools, dense verifiable rewards and counterfactual twin worlds that cap any memoryless agent at 25%, plus procedural "read, take notes, get wiped, act" tasks with randomised tools and schemas. | |
| On Tiny Aya Global (3.35B) we ran a clean ablation with Lexsi AgentTune: two LoRAs trained from base on the identical 4,003 rows (Borderland + DAgger corrections, note-taking tasks, Hermes/Glaive tool calling, GSM8K, Aya multilingual) with identical hyper-parameters. A's agent-task targets include checked derivations or provenance reasons; B's contain the bare actions. GSM8K and Aya targets are identical between arms. | |
| Results on held-out worlds and benchmarks (greedy, 95% CIs): | |
| - Memory across a wipe: after a shared teacher-played prefix through room 3, A writes its own notebook, is wiped, and opens the lock in 90.6% of 64 held-out worlds; B reaches 21.9%, below chance, although it always writes a notebook. Separate room-4 interventions give A 25% with blank notes and 100% following the twin world's teacher-written notes; these do not swap A's own generated notes. On note-taking tasks A reaches 62.5% (base 2.1%) and beats B by +10.4 points (paired CI +3.1 to +18.8). Borderland room success: A 38.4%, B 6.2%, base 0%. | |
| - General ability vs base: both mixed models beat base on BFCL tool calling (base 44.5%, A 54.5%, B 67.0%) and keep multilingual knowledge (Global-MMLU-Lite: base 50.5%, A 54.5%), reversing the collapse of our Borderland-only specialist (BFCL 26.5%). | |
| - Where A loses, and why: our tool-call "reasons" were a provenance template, not working; A learned to over-call and copy raw text (BFCL irrelevance 21%). Derivations help only when they are real computation. Both A and B lost ~15 GSM8K points; shorter answers suggest, but do not establish, imitation of short training solutions as the cause. Neither transfers the note-taking skill to never-trained task types or to an eval-only coding bug-fix wipe task (~0%). | |
| The takeaway for agentic fine-tuning: supervising verified working, not just transcripts, is what makes a small model record the right facts and act on them after a context reset; and it has to be genuine working to help. Everything (environment, data generator, verifiers, models, per-item logs) is released. | |