TL;DR Wrapping coding agent harnesses in a repeated plan-develop-QA loop boosts multi-day autonomous software development performance by an average of 52%. Title: Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement URL: Points 🔁 It leaves existing harnesses like Codex CLI and OpenCode untouched, adding an iterative improvement layer on top 🧭 Splits work into Planner, Developer, and QA Tester roles, carrying verified evidence forward into each new loop 📊 Improved across all 3 benchmarks and 3 model configs, e.g. GameCraft-Bench score jumping from 49.58 to 71.52 💰 Beats plain repeated Vanilla passes even at the same pass count, and is more token-efficient too 🎮 A 70-iteration, multi-day case study built a fully playable FPS game from scratch 🧩 Ablations confirm every piece matters: dropping plan updates, evidence feedback, or warm-start each hurts performance What stands out to me is the philosophy: not just running longer, but iterating smartly by accumulating verified knowledge. #AIAgents# #SoftwareEngineering#

TL;DR Advanced encoders were losing to older models on sparse retrieval for one surprising reason: a vocabulary mismatch. Fixing it with just 20k adaptation steps sets a new BEIR state of the art. Title: Why Advanced Encoders Lag on Sparse Retrieval? The Answer and an Approach to Bridging Vocabulary Gaps URL: Key points 🔤 Raw case-sensitive vocabularies were splitting single meanings across redundant surface forms 🧩 The proposed Vocabulary Transfer method migrates vocabularies in 3 lightweight steps 📊 ModernBERT-VT hits nDCG@10 52.4 on BEIR, a new SOTA 💥 A collapsed RoBERTa-large jumps from BEIR score 1.4 to 51.3 ⚡ Uses under 0.2% of the original pretraining token budget 🧪 Near-optimal performance reached in just 500 MLM steps 🧬 Also works for domain-specific vocabularies like chemistry It's a nice reminder that what looked like an architectural ceiling turned out to be a fixable vocabulary design problem. #InformationRetrieval# #LLM#

TL;DR If the same document gets searched over and over, why re-encode it from scratch for every query? DoPR compresses the document side once, stores it, and reuses it across many queries to speed up LLM reranking. Title: DoPR: Reusable Compressed Document Prefixes for Efficient LLM Reranking URL: 🔑 An attention-based salience score picks the top-K most attended tokens, and each document is compressed once offline into a query-independent prefix of KV states ⚡ Online reranking only has to process the query and score tokens, cutting memory by up to 8.0x and latency by up to 8.04x on long documents 📊 On TREC DL19/DL20 and BEIR, DoPR keeps 97.1-99.5% of NDCG@10, and larger backbones turn out to be more resilient to the compression bottleneck 🔁 In a controlled cross-query reuse test, DoPR hits roughly a 2x speedup at 10 queries per document on DL19, and over 3x at just 5 queries on the long-document TREC-Covid set 🧪 Ablations show attention-guided top-K selection beats first-K, uniform-K, and random-K selection (71.44 NDCG@10) — and it needs zero extra parameters to pull that off 💬 The more a search system relies on a stable, repeatedly-queried document collection, the more this "compute once, reuse many times" idea should pay off #Search# #LLM#

TL;DR A new method runs a large teacher LLM just once to "compile by training" a small, dedicated neural function straight from a natural-language spec. It pushes accuracy on a hard benchmark way up in just minutes. Title: Compile by Training: Turning Natural-Language Specifications into Local Neural Functions URL: 🎯 A teacher mix of GPT-5.4-mini and GPT-5.5 (2:1) auto-generates input-output examples, which are baked into Qwen3-0.6B on the spot via LoRA (rank 64) ⚡ On the hard FuzzyBench-Hard benchmark, LEM accuracy jumps from 0.224 for the fast compiler to 0.836 🕒 Compiling takes about 51 seconds on a B300 GPU, but running teacher data synthesis and training in parallel keeps the GPU from sitting idle 🌐 Validated across real use cases: a QA assistant spanning 4 sites (28 of 30 compiled programs in production) and a natural-language-to-3D-DSL controller (43 of 44 commands succeeded) 📈 Ablations on teacher mixing and data scale show accuracy saturating around 2,400-3,600 unique pairs 💬 Instead of calling a giant model every single time, training it once and carrying home a small function seems like a great fit for medium-complexity tasks that run often #LLM# #LoRA#

TL;DR A new method runs a large teacher LLM just once to "compile by training" a small, dedicated neural function straight from a natural-language spec. It pushes accuracy on a hard benchmark way up in just minutes. Title: Compile by Training: Turning Natural-Language Specifications into Local Neural Functions URL: 🎯 A teacher mix of GPT-5.4-mini and GPT-5.5 (2:1) auto-generates input-output examples, which are baked into Qwen3-0.6B on the spot via LoRA (rank 64) ⚡ On the hard FuzzyBench-Hard benchmark, LEM accuracy jumps from 0.224 for the fast compiler to 0.836 🕒 Compiling takes about 51 seconds on a B300 GPU, but running teacher data synthesis and training in parallel keeps the GPU from sitting idle 🌐 Validated across real use cases: a QA assistant spanning 4 sites (28 of 30 compiled programs in production) and a natural-language-to-3D-DSL controller (43 of 44 commands succeeded) 📈 Ablations on teacher mixing and data scale show accuracy saturating around 2,400-3,600 unique pairs 💬 Instead of calling a giant model every single time, training it once and carrying home a small function seems like a great fit for medium-complexity tasks that run often #LLM# #LoRA#