Register and share your invite link to earn from video plays and referrals.

Search results for TL漫画动漫
TL漫画动漫 community
One keyword maps to one global community path.
Create community
People
Not Found
Tweets including TL漫画动漫
TL;DR: Jetriever shifts agent memory retrieval from similarity to utility. It scores every stored memory with a decision model, so implicit connections get found. Jetriever: Shifting Agent Memory Retrieval from Similarity to Utility Key points ・🥜 A "nut allergy" note matters for a macaron recipe request that never mentions nuts. Similarity search misses it ・🎯 Every memory gets a calibrated "will this help the answer" probability; the whole pool is ranked and the top k go to the answer model ・🧠 Uses Jev, a System One decision model with world knowledge and calibrated probabilities ・📈 On InMind-125 the gold exchange is ranked 1st 84.8% of the time, versus 0.8% for BM25 and 0.0% for MiniLM ・✅ End-to-end accuracy is 76.0%, against 14.4% for the best published memory system ・💸 About $0.008 per query, but cost grows with the number of candidates On MAB (a 60-question sample only) it also rises from 44.3% to 55.4%. But evaluation centers on planted facts, and Jev is accessed through a commercial API. Scaling to large memory stores is the open problem. #AgentMemory# #Retrieval#
Show more
TL;DR: The common claim that dense embeddings beat BM25 in code search depends heavily on how you evaluate. Against a fair BM25 the premium shrinks, and fusing dense with BM25 is the safer choice. Embeddings vs. Grep: Rethinking the Dense Advantage in Code Search Key points ・🧪 Proposes RepoPremium, a repo-level benchmark (core: 1,265 queries over 10 Python repositories) ・📏 Switching BM25 from chunks to whole files shrinks the Cur-Pre premium from +0.121 to +0.034 ・🔤 With generate-then-grep, where an LLM writes identifiers first, BM25 scores 0.613 vs 0.532 for dense ・🤝 RRF fusion of dense and BM25 scores 0.609, above dense alone at 0.550 ・🧭 About 30 routers were tried; none met the preregistered criterion ・🤖 In agent runs, swapping the search tool moves exact-match by only about 0.04 Caveats: mostly 10 Python repositories and localization-only evaluation; other languages use one repository each. #CodeSearch# #RAG#
Show more
TL;DR: MV-IVF is an IVF index built directly over multi-vector (ColBERT-style) data, with point clouds as both documents and centroids. In batched search it averages about 4x the throughput of SVH. MV-IVF: Multi-Vector Retrieval via Multi-Vector Clustering Key points ・🧩 Documents and centroids are both point clouds; routing uses MaxSim ・🔁 With assignments fixed, the center update reduces to ordinary weighted clustering (MV-Lloyd) ・⚡ 1-bit / 4-bit / 8-bit quantization plus query compression, used at different stages ・📦 Batch search groups queries by leaf for better cache reuse ・📈 At 90% Recall-10@10, mean throughput is 2.19x MUVERA, 4.01x SVH, 2.07x MV-Vamana ・⚖️ Single-query latency wins on smaller datasets; MV-Vamana can win on larger ones Early stopping is a heuristic rather than a certified bound, and evaluation covers ColBERTv2 text retrieval only. Still, it shows that clustering multi-vector data directly is practical. #MultiVectorRetrieval# #ColBERT#
Show more
TL;DR: e-means optimizes k-means with EMA and matches Lloyd's 25-epoch loss in about 1.4 to 2.6 epochs. No hyperparameter sweep needed. e-means: outperforming k-means (in one-tenth the epochs) Key points ・⚡ Reaches Lloyd's 25-epoch loss in roughly 1/10 of the epochs (as reported) ・🔁 EMA-updated counts and sums per cluster; centroid read out as S/N ・🌊 Cosine-annealed step size ・🩹 Dead clusters revived by splitting every 5 batches (the biggest effect) ・📐 Decay and batch size set by scaling laws, no sweep ・🚀 DINO10B (10B embeddings) split into 16M clusters in about 2.5 epochs ・🖥️ 9.5-45.5x wall-clock speedup over Flash-KMeans The scaling laws were mostly fitted on DEEP, so accuracy can drop in extreme settings. Clustering large embedding sets just got much cheaper. #MachineLearning# #kmeans#
Show more
TL;DR: Colette: We’re not the central bank of AI, we’re a bank in AI that’s also doing a boatload of buybacks $NVDA
Tl;dr GPT-6.2 is the worst release since the GPT-5 disaster but $200 sub plans now get access to Ultrafast to burn your tokens faster than ever before. Maybe Anthropic will get its $2t valuation back on Thursday after OpenAI helped them with their ipo. What a day wtf
Show more
0
109
2.1K
40
Forward to community
TL;DR: Causal foundation models can report sound intervals per unit, yet become overconfident on group averages. The cause is aggregation that ignores correlated unit errors. Unit-level aggregation breaks coverage in causal foundation models Key points ・🎯 For the full-sample average effect on ACIC 2016, CausalPFN's calibrated 95% interval contained the truth in only 8 of 40 datasets ・🧮 The cause is correlated unit errors: variance is inflated by 1+(k−1)ρ (the design effect), yet units were summed as if independent ・🔁 The fix is SCM-marginalized aggregation: units within a draw share one causal model ・🧠 Approximated with small CNP and latent NP models, also usable as heads on a frozen CausalPFN ・📈 Full-sample coverage across 240 datasets on 6 benchmarks is about 90% (CausalPFN: 53%) ・⚠️ The cost: lower unit-level accuracy on 3 benchmarks and wider intervals If decisions rely on group averages, how uncertainty is aggregated deserves as much scrutiny as the point estimate. #CausalInference# #FoundationModels#
Show more
TL;DR: SO-Arena is a benchmark that measures System One models, which return a decision in a single forward pass, on both capability and reliability. A more capable model is not necessarily a more stable one. SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models Key points ・📚 2,919 tasks across 11 domains, built from 40 public datasets ・⚡ Open-weight System One models answer within 0.3 s per task, tens of times faster than System Two ・📉 Even the best System One trails GPT-6 Luna (79.7) by about 10 points ・🔀 Order stability: does the answer survive reshuffling the options? ・🚫 Null discrimination: can the model tell when the correct option has been removed? ・🔍 The most reliable model, JevAny-27B, ranks 11th of 21 in capability Simple semantic judgments are fine for small models, but multi-step derivation like math benefits greatly from thinking. If you use these for routing or verification, look at stability, not just accuracy. #LLMBenchmark# #SystemOne#
Show more
TL;DR: By generating perspective and bird's-eye-view (BEV) video together, a driving world model can make surrounding vehicles react to the ego vehicle's maneuvers. BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model Key points ・🚗 Prior models took surrounding-vehicle trajectories as external input, fixing them to logged motion ・🗺️ It generates perspective and BEV video jointly, so scene layout becomes part of the prediction rather than an input ・🧠 Two video diffusion transformers use joint attention at every block (about 4B parameters) ・🎲 Independent per-stream noise lets each branch learn to use the other's information ・📉 With ego-action conditioning, ADE drops about 35-36% vs OmniDreams (Plücker) ・👥 In a human study (25 participants), BirdDreams was preferred 57% vs 24% for OmniDreams Caveats: ego trajectories are reconstructed from generated video, and the human study compares only variants given the same road geometry. #WorldModel# #AutonomousDriving#
Show more
TL/DR: the govt is coming for its ton of AI flesh