- llm
- nlp
- reasoning
- rl
- rlhf
- ssm
- theory
- information-theory
- representation-learning
- optimization
- rag
- alignment
- paper-review
- survey
- notes
•
•
•
•
•
•
•
•
•
•
•
•
•
•
-
AlpaServe — Statistical Multiplexing with Model Parallelism for DL Serving
Paper review of AlpaServe (Li et al., OSDI 2023) — the systems paper that argued model parallelism is not just for fitting big models on big clusters but for *statistically multiplexing* bursty multi-model workloads, and showed it can serve requests at up to 10× higher rates or absorb 6× more burstiness at >99% SLO attainment.
-
LeWorldModel — A 15M-Parameter JEPA That Actually Trains End-to-End from Pixels
Paper review of LeWorldModel (Maes, Le Lidec, Scieur, LeCun, Balestriero, 2026) — a Joint-Embedding Predictive Architecture for action-conditioned world modeling whose two-term loss (next-embedding MSE + SIGReg) collapses six tunable hyperparameters into one, trains stably from raw pixels at ~15M parameters on a single GPU, and plans 48× faster than DINO-WM on PushT.
-
FrontierSmith — Manufacturing Open-Ended Coding Problems to Train Better Code Agents
Paper review of FrontierSmith (FrontierCS Team, 2026) — an automated pipeline that mutates closed-ended competitive-programming problems into open-ended ones, filters them with an "idea divergence" metric, and uses the resulting data to substantially boost coding LLMs on FrontierCS and ALE-Bench.
-
ECHO — Terminal Agents Learn World Models for Free
Trend note on ECHO (Shrivastava & Papailiopoulos, 2026) — a one-line change to GRPO for CLI agents that stops masking out terminal-output tokens, adds an environment cross-entropy loss alongside the policy loss, and produces measurable world-model behavior with essentially zero extra compute.
-
MEMENTO: Teaching Reasoning Models to Compress Their Own Thinking
Trend note on MEMENTO (Kontonis et al., 2026, MSR/UW-Madison) — a training recipe that teaches a reasoning model to segment its own chain of thought, compress each segment into a "memento" summary, and evict the original KV cache, yielding 2–3× lower peak KV with near-baseline accuracy.