AI SEARCH · 2026-10-23
AI Search — Paper Daily, 2026-10-23
Community search over short user-generated posts, questions, answers and comments. Key metrics are search relevance (nDCG) and CTR. Focus: semantic matching, query understanding, short-text retrieval, ranking models, and LLM applications in search.
4.75
LEAD STORY ·
PerplexityPerplexity Perplexity releases 0.6B and 9B multi-vector retrievers that encode text, images, and document screenshots: each document token stores a 128-d vector, and scoring sums MaxSim of each query vector against document vectors. Both models distill token-by-token from the same 18B teacher, so they share one vector space and a 9B offline document index can be queried by a 0.6B online encoder. On 72 domain retrieval tasks the 9B model averages nDCG@10 81.3%, 1.6 points above nemotron-3-embed-8b; a 0.6B query with a 9B index scores 79.6%, above 78.0% when both ends use 0.6B. On BrowseComp+ the 9B model lets an agent answer 64.0% correctly with the fewest searches.
4.50
Tsinghua
Renmin UniversityWhen rewriting a query vector with first-round retrieved documents or an LLM-written hypothetical answer (HyDE), the usual approach averages feedback vectors and interpolates with the original query. DIFF argues that each piece of feedback only supports correcting the query along one direction, so other directions should not be pulled back by it: it subtracts the component parallel to the original query to get each feedback's rotation direction and magnitude, constrains only the projection of the correction onto that direction, weights by query–feedback similarity and group agreement, and solves a K×K linear system—without training any model. On 14 BEIR subsets, bge-base with HyDE rises from 0.5492 to 0.5717 mean NDCG@10; with retrieved-document feedback, vector averaging falls to 0.5045 while DIFF scores 0.5522.
3.85
For long-document QA, should you build a vector index then retrieve, or let the model scan the full text segment by segment each time? The authors compare 11 setups on 396 QASPER paper questions and 80 questions from 8 full textbooks, fixing chunking, a top-5 segment budget, and the same answer model. On textbooks, hybrid retrieval plus Jev reranking reaches the highest evidence coverage (94.1%), Jev full-text scan 88.7%, and OpenAI vector retrieval 81.3%: even though full-text scan sees every segment, later filtering drops relevant segments. Higher evidence coverage does not mean better answers; the highest answer accuracy is hybrid retrieval plus OpenAI Decisions reranking (0.892). When queries are many, a reusable vector index costs the least.
3.80
When an ML agent edits code, runs experiments, reads validation scores, and chooses the next step, it keeps changes that happen to look better on that validation set. Walmart defines a feedback rule for this loop: only tell the agent the validation score when the lift over the current best exceeds a noise threshold from paired standard error, and specify when to stop and when the test set may be touched. On a control task constructed so no real lift is possible, an agent that sees exact scores reports a final lift of 0.0157, versus 0.0033 under the rule. On Criteo CTR and MSLR ranking, validation–test gaps shrink and test scores are slightly higher; with exact scores 2 of 3 tasks use all 30 submissions, while the rule stops at 4–12.
3.15
NVIDIAThis NVIDIA tutorial shows how to use KvikIO in Python to read files straight into GPU memory without a CPU bounce buffer (GPUDirect Storage). Each read takes two routing decisions: KvikIO first chooses cuFile or its own POSIX-compatible path, and reads below a threshold (16 KiB in the tutorial) go POSIX; cuFile then chooses the nvidia-fs accelerated path or a CPU-mediated compatibility path. The two choices are controlled by KVIKIO_COMPAT_MODE, KVIKIO_GDS_THRESHOLD, and parallel-split settings in cufile.json. To confirm the accelerated path ran, set CUFILE_STATS_LEVEL=3 and check whether gds_stats counts land in the nvfs or posix column.
3.10
AmazonTorchNeuron is Amazon's open-source PyTorch backend for running models on Trainium. Optional adaptive eager mode batches operators into small graphs that a background compiler compiles, caches, and queues, picking a split strategy by graph size so compilation overlaps with execution of already-compiled work and approaches full-graph compile speed. Artifacts use memory and disk caches, and accelerators on one server share a compile service; on a large cross-accelerator model, time to the first training step falls by about one third. For testing, an AI agent picks 841 representative models by structural features from hundreds of thousands on Hugging Face for tuning and regression.
3.10
NVIDIA NVIDIA announces that DGX Spark officially supports an 8-node cluster, each with 128 GB unified memory (1 TB total), enough to run full GLM 5.3 locally (753B, ~470 GB in NVFP4): about 1,600 token/s single-concurrency prefill and 31 token/s decode. Under switch networking, the RoCEnante collective backend cuts all-reduce latency in tensor-parallel decode, raising the 8-node decode speedup from 3.05× to 4.43×. Pretraining nanochat from scratch on 8 nodes speeds up 7.79× with 97.4% scaling efficiency.
3.10
Linear-attention models keep a fixed-size recurrent state per decoding sequence; the cheaper the state, the more sequences one GPU can serve. Prior work quantizes the state every token and quality drops sharply at 4 bits. PACE quantizes every few tokens and buffers the in-between tokens; it also finds that large values in gated delta state cluster by column, so each column shares one quantization scale; a few heads that accumulate error keep 8 bits. On four hybrid models from 4B to 35B, a 4/8-bit mix matches 8-bit DAMP on the three larger models at about two-thirds the memory; pure 4-bit beats Q-Mamba at equal memory.
3.10
FrontierCS 2 tests whether AI agents can do computer-science research: take a paper and the authors' public code, remove the paper's core contribution, and require the agent to reimplement it in that repo and beat the removed original. Forty-eight tasks span 13 areas including ML systems, HPC, databases, security, and robotics; agents may submit patches anytime for a dev-set score and are judged on a held-out test set against the original. Among 12 agents, Claude Opus 5.5 is best with 58% of runs passing; 16 tasks see no passing run. Runs whose aggregate score beats the original still usually miss some metric or workload.
3.05
Memory systems such as Mem0 and Zep decide what to store before a question arrives; coding agents keep the raw dialogue and let the model search with tools. On one LoCoMo dialogue (419 turns, 152 questions) the authors run four memory methods (full context, BM25, Mem0, Zep) and three coding agents (Pi, Codex, Claude Code) with GPT-4o-mini, GPT-6 Luna, and Claude Opus 5.5, logging cost each time. Mem0 beats full context only under GPT-4o-mini, and building the memory alone costs more than one full-context QA; with Claude Opus 5.5, Pi answers 99% correctly at under half the full-context cost. Each combo is run once; the authors note the study is small.