FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention

Chunyang Li; Dongyang Ma; Dong Yu; Haitao Mi; Jiachen Yu; Jia Li; Miao Peng; Nuo Chen; Qifan Zhang; Tian Liang

REVIEW 4 major objections 4 minor 2 cited by

This paper proposes that a large language model can serve ultra-long contexts by predicting and fetching only the KV chunks the next decoding steps will need, keeping accuracy while cutting GPU KV memory to 13.5% of the baseline.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 11:59 UTC pith:LRCPWTHI

load-bearing objection Useful engineering result with honest self-diagnostics; the abstract overclaims and artifacts are missing, but the core table and failure analysis merit a referee. the 4 major comments →

arxiv 2606.09079 v3 pith:LRCPWTHI submitted 2026-06-08 cs.LG cs.AI

FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention

Yan Wang , Qifan Zhang , Jiachen Yu , Tian Liang , Dongyang Ma , Xiang Hu , Zibo Lin , Chunyang Li

show 7 more authors

Zhichao Wang Miao Peng Nuo Chen Jia Li Yujiu Yang Haitao Mi Dong Yu

This is my paper

classification cs.LG cs.AI

keywords Lookahead Sparse AttentionKV cache compressionmemory indexerlong-context servingdecoupled trainingretrieval-based attentionsparse attentionattention denoising

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that during decoding, an LLM does not need its full KV cache resident in GPU memory: a lightweight Memory Indexer can look ahead and fetch only the compressed historical chunks that upcoming tokens will actually attend to. The indexer is a dual-encoder trained entirely on pre-extracted hidden states and attention-derived labels, never loading the billion-parameter backbone into GPU memory. On three long-context suites (LongBench-v2, LongMemEval, RULER), the resulting system FM-DS-V4 is reported to hold average accuracy at 77.5% (0.6 points above the full-context baseline) while consuming 13.5% of the baseline KV cache footprint, with savings reaching 90% at 500K context. The paper's own diagnostics show the approach collapses on dense multi-range retrieval (MRCR: 76.0 to 48.0) and beyond 2x its training context length, so the central claim is that predictive chunk selection works for a broad class of long-context tasks, not all of them.

Core claim

On the paper's own terms, the central discovery is that a frozen backbone can retain its global memory and reasoning abilities while only a small subset of its compressed historical chunks is physically resident in GPU memory — because a lightweight indexer can predict, τ=64 steps ahead, which chunks the current decoding window will need. The indexer is trained offline as a retrieval model: labels come from cross-layer majority voting over the backbone's own indexer logits (softmax-normalized, top-p filtered at p=0.6, requiring agreement from at least three of 21 layers), and the indexer learns a sigmoid-gated dot-product score with a fixed 0.5 threshold instead of a fixed Top-k. The reporte

What carries the argument

The load-bearing object is the Memory Indexer, a standalone dual-encoder whose only architectural departure from the host model's native Lightning Indexer is a sigmoid activation replacing ReLU, turning raw attention scores into binary classification scores used with a fixed 0.5 threshold. It is placed on three strategic intermediate layers (10, 12, 20) whose binarized decisions are combined by OR-union during decoding. The indexer is trained in complete isolation on pre-computed hidden states and golden labels using focal-loss BCE (γ=2, 3:1 negative sampling), with the compressed indexer keys frozen as the retriever's key side. Inference runs as a tiered selection: the threshold fetches a c

Load-bearing premise

The load-bearing premise is that which compressed chunks the full model will need for the next τ tokens can be predicted by a shallow dot-product scorer on frozen compressed summaries — a distillability assumption that the paper's own diagnostics contradict on dense multi-range retrieval (MRCR) and beyond 2x the training context length.

What would settle it

A direct falsifier is to measure, on real serving logs or a held-out suite, how often decoding with only the last 8K tokens (plus the indexer's fetched chunks) reproduces the outputs of full-context decoding: if the frequency of exact matches falls far below the claimed 90%-of-requests figure, the motivating premise fails. Additionally, any benchmark with the MRCR property — where even the oracle top-50% golden chunks (sorted by true attention density) drop accuracy ~2% below full context — will necessarily defeat any candidate-selection indexer, and a second independent such benchmark would r

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

If the 13.5% average memory footprint holds, long-context serving can raise concurrency (~2.7x) and throughput (~2.8x) on the same GPUs, with per-decode compute dropping to 0.30x the baseline at 1M context.
On tasks where global semantics matter but fine-grained token recall is secondary, the pruned context acts as an attention denoiser: LongBench-v2-L (493K) gains +1.9% over the full-context baseline while using 10% of the memory.
Because the indexer is trained without the backbone, the selection policy can be re-optimized in roughly one GPU hour per configuration, making workload-specific tuning cheap as long as compressed keys and labels are precomputed.
The memory savings grow with context length (up to ~90% at 500K), because the threshold-based recall set grows sublinearly in sequence length, unlike the linear full-KV cache.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The paper's §1 assertion that over 90% of real requests with contexts longer than 64K tokens can be answered from the last 8K tokens is unverified; if it holds, serving could use a two-tier policy (local window first, indexer only for the residual) and reserve full-context attention for the minority of requests that need it.
The MRCR failure suggests a testable refinement: chunk-level selection is the bottleneck, not the classifier — since even loading the oracle top-50% golden chunks loses ~2% on that benchmark, a late-interaction or token-level scorer may be required for dense global memory tasks.
The observed 2x length-generalization ceiling implies the indexer's pointwise scoring inherits the positional-embedding distribution of its training contexts; length-annotated training mixtures or length-adaptive thresholds are plausible extensions, but remain beyond the paper's claims.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Useful engineering result with honest self-diagnostics; the abstract overclaims and artifacts are missing, but the core table and failure analysis merit a referee.

read the letter

The genuinely new thing here is the decoupled dual-encoder indexer: a lightweight query encoder trained against frozen compressed indexer keys, with labels extracted offline from the frozen backbone, so the entire indexer trains in about an H20 GPU-hour without loading the model. That is a practical recipe people will want to steal. The paper also does an unusually honest job of reporting its own failures—the Project Status block, the admitted untested hyperparameters, and a §3.3 that documents the MRCR collapse (76 to 48), the oracle result that even true top-50% golden chunks lose ~2% on that benchmark, the context-free leakage, and the length ceiling at 2× training context.

On the merits, Table 1 is coherent: 13.5% KV memory at rough parity or slight gains on LongBench-v2, LongMemEval, and RULER. I buy that as a useful engineering result for DeepSeek-V4-style compressed-attention serving, and the controls (Recency Only, Random 10%) strengthen the claim that the indexer is doing real selection.

The soft spots are real and mostly about scope, not about the core table. The abstract overclaims: "consistently preserving" is contradicted by Table 1 itself (RULER-128K 94.3→93.2, RULER-256K 90.5→88.2), and the 1M-context numbers (0.30× compute, 3.73→0.37 GB) appear nowhere in Table 1 and sit right at or beyond the paper's own 2× length ceiling—§3.3.3 says beyond 2× training length selection degenerates to near-random. The "over 90% of real requests" claim in §1 has no supporting data. And the parity is partly by construction: the labels are distilled from the same frozen backbone's softmax/Top-p/majority-vote, so the indexer is learning to imitate the teacher's selection, which is legitimate but makes the "attention denoiser" framing speculative. The +0.6% average actively hides the MRCR collapse by excluding it.

Two more moderate concerns. No code, data, or weights; the GPU-memory numbers rest on sglang deployment logs that are not shown, so I can't verify anything independently. And the headline "13.5%" mixes contexts from 46K to 512K, so the average is fine but it flatters the short end.

The core claim is still plausible and the paper's own diagnostics are honest enough that I'd trust the Table 1 numbers more than the abstract. This is a paper for practitioners of long-context inference and for researchers working on learned sparse attention or KV selection. It deserves a serious referee, not a desk reject, but the required revision is real: scope the claims to the tested suites and lengths, ship the indexer weights and evaluation code, and show the deployment measurements. If the authors do that, this becomes a solid systems result.

Referee Report

4 major / 4 minor

Summary. The paper proposes Lookahead Sparse Attention (LSA), an inference-time mechanism for DeepSeek-V4-Flash that uses a lightweight Memory Indexer to predict, every τ decoding steps, which compressed CSA chunks should be loaded into GPU memory. The indexer is a dual-encoder trained offline on pseudo-labels extracted from the frozen backbone via softmax normalization, Top-p filtering, and cross-layer majority voting (Eqs. 7–11). Experiments on LongBench-v2, LongMemEval, and RULER report that FM-DS-V4 preserves or slightly improves average accuracy (+0.6%) while reducing average KV-cache memory to 13.5% of the full-context baseline. The paper also claims substantial 1M-context efficiency gains (0.30× compute, 3.73→0.37 GB, 2.8× throughput, 2.7× concurrency). The authors transparently document three failure regimes: context-independent overhead leakage, a severe MRCR accuracy collapse (76.0→48.0), and a length-generalization ceiling at 2× the training context length.

Significance. If the core claim is validated, this is a practically valuable contribution: reducing KV-cache memory by roughly 7× while maintaining accuracy on three standard long-context benchmarks is a meaningful step for long-context serving, and the backbone-free decoupled training pipeline (reported to finish in one H20 GPU hour) is a strong reproducibility feature. The paper also deserves credit for honestly reporting counterexamples and failure modes instead of hiding them. However, the evidence as presented supports the efficiency/accuracy claim only for the three named suites at context lengths up to 512K; the 1M numbers are unmeasured, and the MRCR and length-generalization results show fundamental limits to the indexer-based selection premise. The contribution is real but narrower than the abstract and introduction claim.

major comments (4)

[Abstract; §3.3.3] The abstract’s 1M-context numbers (0.30× compute, 3.73→0.37 GB, 2.8× throughput, 2.7× concurrency) are not present in Table 1 or any experimental section. §3.3.3 states that the released indexer was trained only on contexts up to 512K, that empirical validation at greater scales remains untested, and that retrieval discriminability would decay irreversibly beyond 1M. Thus the headline 1M claims are unsupported and should be removed or replaced with measured results.
[Table 1; §3.2; §3.3.2] The claim of “consistently preserving or slightly elevating” accuracy is contradicted by the paper’s own Table 1: RULER-128K drops from 94.3 to 93.2 and RULER-256K from 90.5 to 88.2, both outside a small noise margin. The +0.6% average is driven by LongBench and LongMemEval. Moreover, §3.3.2 reports MRCR accuracy falling from 76.0 to 48.0, and even an oracle with the true top-50% golden chunks loses ~2%, showing a task class where no sparse selection can preserve accuracy. The headline should be scoped to the tested suites and explicitly acknowledge the MRCR failure.
[§2.2, Eqs. (7)–(11); §3.3.2] The positive labels Y_t^+ are constructed from the frozen backbone’s own softmax/Top-p/majority-vote decisions, so the Memory Indexer is trained to imitate the teacher’s retrieval behavior. Accuracy parity at high recall is therefore partly by construction and does not establish that the indexer independently identifies “golden” chunks. The MRCR oracle experiment (top-50% golden chunks still lose ~2%) demonstrates that the teacher’s attention-derived labels are themselves insufficient for dense global-memory tasks. The paper should present this as a bound on the approach, not merely future work.
[§3.3.3] The paper states the indexer safely generalizes only up to 2× the training context length and collapses into near-random selection beyond that. With training at 512K, that places 1M in an untested boundary zone, not a validated regime. The conclusion’s claim that the results justify confidence in “ultra-long-context intelligence” and the abstract’s 1M figures are therefore not supported by the evidence. The central claim should be explicitly limited to ≤512K or to 1M with the stated hypothesis clearly separated from experimental results.

minor comments (4)

[§1] The claim that “over 90% of user requests with contexts longer than 64K tokens can be accurately resolved using only the last 8K tokens” is presented without evidence or citation. Please provide the inference-log analysis or mark this as a motivating hypothesis.
[§2.4, ] The sentence “the per-sample weight wt,s computed by the–weighted-loss scheduler” appears garbled; define wt,s explicitly. Also, the Note on Hyperparameter Selection admits τ=64 and threshold=0.5 were never ablated; please report sensitivity results or state these as open parameters in the main text.
[Throughout] There are several typos and informal expressions, e.g., “an golden label” (§2.2), “threadbare” (§3.2), and “the-2.3” in the limitations section. A proofread pass is needed.
[Front matter] The inserted “Project Status” paragraph and the call for collaborators/contact email are not appropriate for a formal technical report. Remove or move to an acknowledgments section.

Circularity Check

0 steps flagged

No formal circularity: the indexer is trained on labels extracted from the frozen backbone, but downstream accuracy is externally benchmarked and the paper's own negative results show the claim is not forced by construction.

full rationale

The training target Y_t^+ (Eq. 11) is computed from the frozen DS-V4-Flash backbone's own softmax/Top-p/majority-vote scores (Eqs. 7-10), and the indexer is optimized to reproduce these labels (Eqs. 12-15). This is a self-distillation / imitation setup, not a circular derivation: the reported accuracy on LongBench-v2, LongMemEval, and RULER is measured against external task labels, and the paper reports concrete failure cases (MRCR 76.0→48.0; the oracle top-50% simulation still loses ~2%; RULER-128K 94.3→93.2; RULER-256K 90.5→88.2; length-generalization collapse beyond 2x training length in Secs. 3.3.2-3.3.3) that falsify any claim that preserved accuracy is guaranteed by construction. The 'Note on Hyperparameter Selection' also explicitly says tau=64 and the 0.5 threshold were chosen from exploratory runs and not ablated. The only self-citation (ref. [4], LongMemEval, co-authored by Dong Yu) is a benchmark reference and is not load-bearing. The abstract's 1M-context efficiency numbers are extrapolations beyond the 512K training/validation boundary and are not measured, which is an overclaim but not a circular step. No enumerated circularity pattern (self-definitional, fitted-input-called-prediction, self-citation-load-bearing, uniqueness-import, ansatz-via-citation, renaming-known-result) is exhibited.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 1 invented entities

The efficiency and parity claims rest on: (i) hand-chosen hyperparameters (tau=64, 0.5 threshold, p=0.6, theta=3, gamma=2, 3:1 negatives, r=2048, layers {10,12,20}) that are only partially ablated; (ii) treating the frozen backbone's own selections as ground truth, so the indexer is a distilled imitation rather than an independent predictor; (iii) a coarse similarity assumption (Eq. 4) that the paper itself shows fails on MRCR and beyond 2x training length; and (iv) deployment measurements (sglang logs, throughput/concurrency) asserted but not shown. The unverified '90% of long requests need only last 8K' premise also underlies the framing.

free parameters (8)

decoding trigger interval tau = 64
Defines the lookahead window; chosen in exploratory runs; paper admits it is untested across alternatives (§2.4 note).
classification threshold (Sigmoid cutoff) = 0.5
Controls which chunks are fetched (Eq. 5); admitted untested across values (§2.4 note).
Top-p nucleus threshold in label filtering = 0.6
Used in Eq. 8 to denoise golden labels; set empirically with no ablation reported.
majority-vote threshold theta = 3
A chunk becomes a golden label only if at least 3 layers vote for it (Eq. 10); no ablation reported.
focal loss focusing parameter gamma = 2
Down-weights easy negatives in Eq. 15; standard value, not swept.
negative sampling ratio = 3:1
Three negatives per positive; chosen to handle class imbalance; no ablation reported.
query low-rank projection dimension r = 2048
Increases query-encoder capacity (q_lora_rank); chosen for the R-series configuration (§2.4).
indexer layer geometry = {10, 12, 20}
Selected via a 500-run Pareto sweep; the paper notes a finer layer-wise ablation is still desirable (§2.4).

axioms (5)

domain assumption The cited DeepSeek-V4-Flash backbone (HCA 128:1, CSA layers, native Lightning Indexer, 21 CSA layers) exists and behaves as referenced [1].
Every baseline and label-generation step depends on this external model; only a HuggingFace URL is given and nothing about it is reproduced in this paper.
domain assumption The frozen backbone's own Softmax/Top-p/majority-vote selections (Eqs. 7-11) are a correct and sufficient ground truth for which chunks a future token needs.
The indexer inherits any error in the teacher's choice; §3.3.2's MRCR collapse shows this ground truth is insufficient for dense global memory.
domain assumption A coarse dot-product of the current query hidden state with compressed chunk summaries, fused across heads with ReLU and gated by sigmoid (Eq. 4), is a sufficient statistic for future context demand.
This is the paper's core modeling hypothesis; it fails on MRCR and beyond 2x training length per §§3.3.2-3.3.3.
domain assumption Over 90% of user requests with contexts longer than 64K tokens can be resolved using only the last 8K tokens.
Motivates the whole asymmetry in §1, but no data or analysis is shown to support this empirical premise.
standard math Routine calculus and probability used in the BCE and Focal Loss objectives (Eqs. 12-15).
Standard derivations; no new mathematics is introduced.

invented entities (1)

Memory Indexer (Lookahead Sparse Attention block selector) no independent evidence
purpose: Dual-encoder scorer that predicts which compressed KV entries a future tau-token window needs and fetches them from CPU to GPU (Eqs. 1-5).
A software component proposed and evaluated only on the paper's own benchmark runs; no weights, code, or external handle is provided, so there is no falsifiable handle outside this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 11463 in / 20896 out tokens · 206218 ms · 2026-08-02T11:59:55.578410+00:00 · methodology

0 comments

read the original abstract

Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving. In this report, we propose \textbf{Lookahead Sparse Attention (LSA)}, a novel inference paradigm powered by a Neural Memory Indexer built upon the DeepSeek-V4 architecture. Rather than passively attending to all historical tokens, LSA proactively predicts future context demands and preserves only the query-critical KV chunks in the GPU memory. Crucially, we instantiate this architecture via a \textbf{backbone-free decoupled training} strategy. By formulating the indexer as a standard dual-encoder architecture, we train it independently using standard retrieval training frameworks without ever loading the massive backbone model into GPU memory. We demonstrate that this ``less is more'' paradigm significantly maximizes serving efficiency while acting as an effective attention denoiser in tasks that rely on long-term global memory. Across primary long-context evaluation suites (e.g., LongBench-v2, LongMemEval, and RULER), \texttt{FM-DS-V4} compresses the average physical KV cache footprint down to merely 13.5\% of the full-context baseline, while consistently preserving or slightly elevating downstream accuracy (+0.6\% absolute margin on average). At 1M context, per-decode-token compute drops to 0.30$\times$ of the baseline and GPU KV cache shrinks by 90\% (3.73$\to$0.37 GB), translating into \textbf{2.8$\times$ aggregate throughput and 2.7$\times$ concurrency gains} in PD-disaggregated serving on 8$\times$H20 GPUs.

Review history (2 revisions) →

discussion (0)

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Semantic-aware and Self-improving Program Reduction via Agentic Large Language Models
cs.SE 2026-07 conditional novelty 7.0

An agentic LLM reducer plus a reflector that learns reusable strategies produces 36–39% smaller minimized bug programs than the best prior reducers on 90 C/Rust/JS benchmarks.
Variational-Ising-Attention (VIA):TailoredAttentionMattersfor Science
cs.LG 2026-07 conditional novelty 6.0

Variational-Ising-Attention with backbone-deployed low-rank couplings substantially beats softmax on USPTO-50k reaction-center exact-match, especially for multi-center cases.

Reference graph

Works this paper leans on

6 extracted references · cited by 2 Pith papers

[1]

Deepseek-v4: Towards highly efficient million-token context intelligence

DeepSeek-AI. Deepseek-v4: Towards highly efficient million-token context intelligence. Technical report, DeepSeek- AI, 2026. Technical Report. Available athttps://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/ DeepSeek_V4.pdf

2026
[2]

Qwen3.5: Extending the frontier of open large language models

Qwen Team. Qwen3.5: Extending the frontier of open large language models. Qwen AI Blog, 2026.https: //qwen.ai/blog?id=qwen3.5

2026
[3]

Longbench v2: Towards deeper understanding and reasoning on realistic long-context multitasks, 2025

Yushi Bai, Shangqing Tu, Jiajie Zhang, Hao Peng, Xiaozhi Wang, Xin Lv, Shulin Cao, Jiazheng Xu, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. Longbench v2: Towards deeper understanding and reasoning on realistic long-context multitasks, 2025

2025
[4]

Longmemeval: Benchmarking chat assistants on long-term interactive memory, 2025

Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Longmemeval: Benchmarking chat assistants on long-term interactive memory, 2025. 10

2025
[5]

Ruler: What’s the real context size of your long-context language models?, 2024

Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What’s the real context size of your long-context language models?, 2024

2024
[6]

Michelangelo: Long context evaluations beyond haystacks via latent structure queries, 2024

Kiran Vodrahalli, Santiago Ontanon, Nilesh Tripuraneni, Kelvin Xu, Sanil Jain, Rakesh Shivanna, Jeffrey Hui, Nishanth Dikkala, Mehran Kazemi, Bahare Fatemi, Rohan Anil, Ethan Dyer, Siamak Shakeri, Roopali Vij, Harsh Mehta, Vinay Ramasesh, Quoc Le, Ed Chi, Yifeng Lu, Orhan Firat, Angeliki Lazaridou, Jean-Baptiste Lespiau, Nithya Attaluri, and Kate Olszewsk...

2024

[1] [1]

Deepseek-v4: Towards highly efficient million-token context intelligence

DeepSeek-AI. Deepseek-v4: Towards highly efficient million-token context intelligence. Technical report, DeepSeek- AI, 2026. Technical Report. Available athttps://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/ DeepSeek_V4.pdf

2026

[2] [2]

Qwen3.5: Extending the frontier of open large language models

Qwen Team. Qwen3.5: Extending the frontier of open large language models. Qwen AI Blog, 2026.https: //qwen.ai/blog?id=qwen3.5

2026

[3] [3]

Longbench v2: Towards deeper understanding and reasoning on realistic long-context multitasks, 2025

Yushi Bai, Shangqing Tu, Jiajie Zhang, Hao Peng, Xiaozhi Wang, Xin Lv, Shulin Cao, Jiazheng Xu, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. Longbench v2: Towards deeper understanding and reasoning on realistic long-context multitasks, 2025

2025

[4] [4]

Longmemeval: Benchmarking chat assistants on long-term interactive memory, 2025

Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Longmemeval: Benchmarking chat assistants on long-term interactive memory, 2025. 10

2025

[5] [5]

Ruler: What’s the real context size of your long-context language models?, 2024

Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. Ruler: What’s the real context size of your long-context language models?, 2024

2024

[6] [6]

Michelangelo: Long context evaluations beyond haystacks via latent structure queries, 2024

Kiran Vodrahalli, Santiago Ontanon, Nilesh Tripuraneni, Kelvin Xu, Sanil Jain, Rakesh Shivanna, Jeffrey Hui, Nishanth Dikkala, Mehran Kazemi, Bahare Fatemi, Rohan Anil, Ethan Dyer, Siamak Shakeri, Roopali Vij, Harsh Mehta, Vinay Ramasesh, Quoc Le, Ed Chi, Yifeng Lu, Orhan Firat, Angeliki Lazaridou, Jean-Baptiste Lespiau, Nithya Attaluri, and Kate Olszewsk...

2024