Princeton University · Research Talk

Auto
Research

Toward autonomous superintelligence: principled architectures and agentic RL that automate AI research.

Press → or Space to begin  ·  ? for help

Research agenda

Toward autonomous superintelligence

The route is AutoResearch: automating AI research itself, so improvements compound. That goal sets three pillars.

I · Agentic RL Scaled to long-horizon tasks Objectives derived, not tuned; training systems that scale. FlashREINFORCE · RPG · GPM & GPO
II · Model Architectures Long context & continual learning Structured state that holds long rollouts and keeps learning. TPA · GRAPE · RLT
III · AutoResearch Automating AI research Agents and harnesses that coevolve with the model. Agora

NeurIPS Spotlight

The thesis

Superintelligence as a loop, not a leap

Architectures long context · continual learning Agentic RL long-horizon training Research agents propose · verify · ship Better models and better methods improvements compound: the loop is AutoResearch
  • 01The goal is AutoResearch: automating AI research itself, so progress compounds.
  • 02Architectures and agentic RL are not separate interests; they are what the loop demands.

Selected works · I

Agentic RL and AutoResearch

NVIDIA (contributors listed alphabetically, including Yifan Zhang)
NVIDIA Technical Report · arXiv:2610.38095
Yifan Zhang, Yunheng Zou, Shaokun Zhang, Jian Hu, Hao Zhang, Binfeng Xu, Jan Kautz, Yi Dong
NeurIPS 2026 AutoMLR Workshop Oral · arXiv:2609.18094
Jian Hu*, Yifan Zhang*, Hao Zhang*, Binfeng Xu*, Shaokun Zhang, Hongqing Peng, Zhiding Yu, Pavlo Molchanov, Jan Kautz, Yi Dong
NVIDIA Technical Report · arXiv:2611.18627
Yifan Zhang and NVIDIA Research
Technical report · August 2026
Jian Hu, Huiying Li, Hao Zhang, Binfeng Xu, Yifan Zhang, Shaokun Zhang, Hemil Desai, Michael Demoret, Pavlo Molchanov, Jan Kautz, Yi Dong
NVIDIA Technical Report · arXiv:2607.21653
Yifan Zhang*, Yifeng Liu*, Huizhuo Yuan, Yang Yuan, Quanquan Gu, Andrew C Yao
ICLR 2026
Yifan Zhang*, Ge Zhang*, Yue Wu*, Kangping Xu, Quanquan Gu
ICML 2025

* Equal contribution

Selected works · II

Pretraining and model architectures

Yifan Zhang, Jichen Feng, Shihan Qin, Sun-Yuan Kung, Mengdi Wang
arXiv:2610.07591
Yifan Zhang, Steve Ta, Jasper Zhang, Jichen Feng, Shuzhen Li, Yongxin Zhang, Yifeng Liu, Huizhuo Yuan, Mengdi Wang†, Quanquan Gu†, Andrew Chi-Chih Yao†
ByteDance Seed Technical Report · arXiv:2608.27763
Yifan Zhang, Yifeng Liu, Mengdi Wang, Quanquan Gu
arXiv:2601.00417
Yifan Zhang, Zixiang Chen, Yifeng Liu, Zhen Qin, Huizhuo Yuan, Kangping Xu, Yang Yuan, Quanquan Gu, Andrew C Yao
ICLR 2026
Yifan Zhang*, Yifeng Liu*, Huizhuo Yuan, Zhen Qin, Yang Yuan, Quanquan Gu, Andrew C Yao
NeurIPS 2025 Spotlight

* Equal contribution · † Corresponding author

The question

What would it take to automate AI research itself?

  • 01RL that scales: today's policy gradients are a brittle patchwork of heuristics.
  • 02Models that hold long horizons: today's attention is memory-bound, and learning stops after pretraining.
  • 03A loop that closes: model and harness must coevolve for improvements to compound.

My research makes each piece principled, and ships the result.

Pillar I

Agentic RL

Derive the objective, then scale agentic RL to long horizons.

Pillar I · two routes

REINFORCE vs. Online Mirror Descent

Route 1 · REINFORCE Follow the reward gradient Policy gradient on the scalar reward, with no critic, no KL term and no ratio clipping. Stale rollouts are corrected by importance ratios and a trajectory gate.
$\nabla_\theta J=\mathbb{E}\big[A\,\nabla_\theta\log\pi_\theta\big]$
FlashREINFORCE
Route 2 · Online Mirror Descent Take a KL-proximal step Each update trades the learning signal against a KL penalty to the current policy. The KL direction, its form and the loss estimator are design choices.
$\pi_{t+1}(y\mid x)\propto\pi_t(y\mid x)\exp\!\big(\eta\,s(y\mid x)\big)$
RPG · GPM & GPO · Score Centering

FlashREINFORCE takes the first route; RPG, GPO and score centering take the second.

Pillar I · FlashREINFORCE

Single-rollout asynchronous REINFORCE

  • 01One fresh batch. One rollout per prompt. Center rewards across independent prompts, then take one update and discard the batch. No learned critic.
  • 02Ratio + trajectory gate. Use recorded sampler probabilities for token importance ratios; screen complete trajectories by a divergence statistic. No ratio clipping.
  • 03Sample-mean optimization. Average tokens within each trajectory, then average trajectories, so long failures do not dominate by length.
FlashREINFORCE at a glance: a one-batch update with a mean baseline, a sequence trust region that admits complete trajectories, and a sample-mean loss in place of a token-mean loss.

FlashREINFORCE · arXiv:2611.18627 ↗ · Full FlashREINFORCE talk ↗
mᵢ admits a complete trajectory. Token ratios correct actions at stored histories, not the history distribution.

Pillar I · FlashREINFORCE · experiments

Reasoning, tools, and long-horizon agents

SettingResultComparison / protocol
Long-CoT · R1-Distill 1.5B21.7 → 33.7AIME24/25 mean; 6,000 updates, lag ≈4
Qwen2.5-Math-1.5B38.0 vs 36.3Five-benchmark mean; 256k vs 512k rollouts
Python tools · Qwen3-30B-A3B+6.8 pointsMatched-budget GRPO comparison
ALFWorld · Qwen2.5-7B98.3% / 96.5%Seen / unseen; 12.8k training trajectories

One rollout per prompt removes sibling-response synchronization and broadens prompt coverage.

FlashREINFORCE · arXiv:2611.18627 ↗ · Results from the public technical report; benchmark means are percentages.

Pillar I · FlashREINFORCE · scaling

Stable on a 30B MoE at policy lag 8

  • 01+6.8 points over GRPO on Qwen3-30B-A3B at the same 102.4k-rollout budget, at policy lag ≈8 against ≈1.
  • 02Holds its accuracy. With routing replay, the three-benchmark mean stays at 63.7–66.8 over updates 1,000–1,450 and ends at 66.4.
  • 03GRPO drops sharply twice before recovering to 60.3 at step 800.
Qwen3-30B-A3B Python-tool accuracy over policy updates on the three-benchmark mean, AIME 2025, AMC 2023 and Minerva: three FlashREINFORCE configurations at lag eight stay stable, while GRPO at lag one drops sharply twice.

FlashREINFORCE · arXiv:2611.18627 ↗ · Python-tool accuracy, avg@4; the mean weights AIME 2025, AMC 2023 and Minerva equally.

Pillar I · Prefill–Decode Kernel Mismatch · Yifan Zhang and NVIDIA Research

Same weights, two policies

  • 01Prefill–decode kernel mismatch. The sampler’s cached decoding and the trainer’s parallel scoring assign different probabilities at identical weights, with no parameter lag.
  • 02Measured gap. On a matched 8,192-token prompt, vLLM and PyTorch differ by 0.093 in sampled-token log-probability (95th percentile) for Qwen3-4B, and by 0.044 for the hybrid Qwen3.5-4B.
  • 03Attention vs. recurrence. Batch-invariant arithmetic aligns attention, whose cache stores inputs. Recurrent layers round the state at different points, so it does not align them.
Where the sampler and the trainer store and round intermediate results. In softmax attention both paths share per-token KV entries. In linear attention and SSMs the chunkwise trainer forms the state only at chunk boundaries, while a tokenwise decoder rounds and stores it after every token.

Prefill–Decode Kernel Mismatch · project page ↗ · Figure: where the sampler and the trainer store and round intermediate results for the same response.

Pillar I · Prefill–Decode Kernel Mismatch · Yifan Zhang and NVIDIA Research

The recurrent schedule alone opens the gap

Qwen3.5-4B with only its Gated DeltaNet cores switched to the FP32 tokenwise rule, everything else fixed.

  • 01On its own samples. Sampled-token |Δ log p| is 0.061 at the 95th percentile, and the top-1 token flips at 0.43% of positions.
  • 02Larger off-policy. Books reach 0.089 and 1.22%; Wikipedia in four languages 0.086–0.096 and 1.07–1.28%.
  • 03Gradients move too. The two schedules’ vector–Jacobian products for queries, keys, values, gates and β differ by 0.3–1.7%.
Sampled-token log-probability gap quantiles, top-1 flip rate and mean KL between tokenwise and chunkwise Gated DeltaNet schedules, on the model's own samples and on off-policy text in several languages.

Prefill–Decode Kernel Mismatch · project page ↗ · Figure 4 of the paper: p50/p95/p99 |Δ log p|, top-1 flips and mean KL per corpus.

Pillar I · Prefill–Decode Kernel Mismatch · Yifan Zhang and NVIDIA Research

Chunk-aligned decoding

  • 01Match the schedule, not just the precision. The gap comes from where the recurrent state is rounded, so an FP32 state does not remove it.
  • 02Keep what the trainer reads. The decoder keeps the last chunk-boundary state and the open chunk’s inputs, so each token comes from the trainer’s chunk kernel.
  • 03Bit-exact in vLLM. Cached decoding reproduces the full prefill bit for bit on all 24 traces.
What must match for one decode step inside a chunk: the state shape, the kernel shape and the update rule. The tokenwise decoder differs from the chunkwise trainer on all three; chunk-aligned decoding matches all three.

Prefill–Decode Kernel Mismatch · project page ↗ · Figure: what chunk-aligned decoding aligns at one decode step inside a chunk.

Pillar I · Prefill–Decode Kernel Mismatch · Yifan Zhang and NVIDIA Research

Aligned kernels keep accuracy and remove the kernel term

AIME 2024, 2025 and 2026 accuracy of base Qwen3.5-4B under chunk-aligned and tokenwise sampling, with paired differences and 95% intervals.
Chunk-aligned sampling, base Qwen3.5-4B. AIME avg@32 macro accuracy is 0.506, against 0.499 tokenwise; every 95% interval of the difference includes zero.
Asynchronous RL on R1-Distill-1.5B at lag eight: training accuracy, policy entropy and sampled forward KL for aligned and ordinary kernels.
Batch-invariant kernels in asynchronous RL, R1-Distill-1.5B at lag eight. Sampled forward KL averages 3.7×10⁻⁴ aligned and 5.4×10⁻⁴ ordinary over the first 3,600 updates.
  • →Bit-exact logits. Over about 40 million response tokens, the aligned arm’s trainer-to-sampler log-ratio has mean −3.4×10⁻⁷ and largest magnitude 6.2×10⁻⁶; the ordinary twin reaches 0.085 at the 95th percentile.

Prefill–Decode Kernel Mismatch · project page ↗ · Figures 6 and 7 of the paper.

The problem

Policy gradients for LLMs are a zoo

  • 01RLHF and RLVR now drive frontier reasoning.
  • 02But GRPO / PPO are a patchwork of clipping and KL hacks — unstable, hard to reason about.

Start from the KL-regularized objective — and be exact.

ICLR 2026

KL-regularized policy gradient, done right

A unified design space for KL-regularized policy-gradient algorithms — with the correct gradient and stop-gradient treatment.

Impact Adopted in DeepSeek-V3.2.
  • →More stable training; recovers and repairs popular methods as special cases.
The RPG framework: inputs, the core engine that constructs and optimizes the KL-regularized objective, and the three design choices: KL direction, KL form, and loss estimator.

Pillar I · RPG · ICLR 2026

The exact gradient for every KL variant

RPG-REINFORCE with RPG-Style Clip gains up to +6 points over DAPO on AIME24 and AIME25.

  • 01One derivation covers forward and reverse KL, normalized and unnormalized; the widely used k3 penalty is exactly the unnormalized KL.
  • 02Stop-gradient done right. Conditions under which REINFORCE-style losses match fully differentiable surrogates in gradient.
  • 03Repairs GRPO. Its KL term has an off-policy importance-weighting mismatch; RPG corrects it.
  • 04RPG-Style Clip. Clipped importance sampling keeps off-policy training stable at scale.
AIME24 and AIME25 mean@32 accuracy over 400 training steps: the RPG and RPG-REINFORCE variants end above DAPO and well above GRPO.

Adopted in DeepSeek-V3.2 · Curves: AIME24 and AIME25 mean@32 accuracy during training, 4k context length.

Pillar I · GPM & GPO · ICML 2025

Beyond Bradley–Terry preferences

Preference embedding places each response in a latent space, so the model can represent cyclic, intransitive human preferences that a scalar reward cannot.

  • →GPM embeds each response once (linear queries, not $\mathcal{O}(K^2)$) and consistently outperforms the Bradley–Terry reward model on RewardBench.
  • →GPO generalizes reward-based RLHF, and each iteration is an online mirror descent step: $\pi_{t+1}(y\mid x)\propto\pi_t(y\mid x)\exp\!\big(\eta\,s(y\succ\pi_t\mid x)\big)$.
  • →At the token level with $\beta=0$, GPO reduces to score centering.
Bradley–Terry reward model, pairwise preference models, and the General Preference embedding Model, which embeds each response once and scores pairs with a skew-symmetric operator.

Pillar I · GPM & GPO · experiments

GPM models what Bradley–Terry cannot

SettingBT reward modelGPMGain
Cyclic preferences · dataset No. 162.4%100.0%+37.6
Cyclic preferences · dataset No. 350.0%100.0%+50.0
RewardBench average · Gemma-2B-it74.8582.29+7.44
RewardBench average · Llama-3.1-8B-Instruct90.5691.90+1.34
  • →Cyclic preferences: GPM reaches 100% where the BT model stays near the 50% random-guess level.
  • →RewardBench: the largest gains are in Chat and Chat-Hard, 67.32 → 79.61 and 63.37 → 75.66 with Gemma-2B-it.
  • →GPO: models aligned with GPO reach higher AlpacaEval 2.0 win rates than BT-based alignment.

GPM & GPO · arXiv:2410.02197 ↗ · ICML 2025 · Cyclic rows are accuracy; GPM embedding dimension is 6 for Gemma-2B-it and 8 for Llama-3.1-8B-Instruct.

Pillar I · online mirror descent

RPG, GPO and score centering share one objective

OMD step ($\eta=1/\beta$): $\;\pi^{+}=\arg\max_{\pi}\;\mathbb{E}_{z\sim\pi}\,f(h,z)-\beta\,\mathrm{KL}\big(\pi\,\|\,\mu(\cdot\mid h)\big)\;\;\Longrightarrow\;\;\beta\log\dfrac{\pi^{+}(z\mid h)}{\mu(z\mid h)}=f(h,z)-c_f(h)$
RPG · response level Maximize the regularized return, with $f=R$ and $w_\theta=\pi_\theta/\mu$:
$\mathcal{L}_{\mathrm{RPG}}=\mathbb{E}_{\mu}\big[-w_\theta R+\beta\,(w_\theta\log w_\theta-w_\theta)\big]$
GPO · response level Regress the log-ratio onto the preference score, with $f=u(x,y)=\mathbb{E}_{y'\sim\mu}\,s(y\succ y'\mid x)$:
$\mathcal{L}_{\mathrm{GPO}}=\mathbb{E}_{\mu}\Big[\log\dfrac{\pi_\theta}{\mu}-\dfrac{u-c_u}{\beta}\Big]^{2}$
Score centering · token level 1. Take the same step at every token, with $f=A^{\mu}(s,a)$, the sampler advantage:
$\beta\,\ell_{\theta,u}=A^{\mu}(s_u,a_u)-c(s_u),\qquad \ell_{\theta,u}=\log\dfrac{\pi_\theta(a_u\mid s_u)}{\mu(a_u\mid s_u)}$
2. Fit this equation by least squares on the sampler’s tokens. The unknown $c(s)$ is replaced by $k_\theta(s)=\mathrm{KL}(\mu\,\|\,\pi_\theta)$, which gives the residual $\delta_u$ zero mean under $\mu$:
$\delta_u=\dfrac{A^{\mu}(s_u,a_u)}{\beta}-\ell_{\theta,u}-k_\theta(s_u),\qquad \mathcal{L}=\dfrac{\beta}{2}\,\mathbb{E}_{\mu}\textstyle\sum_u\delta_u^{2}$
3. Differentiate: $\nabla_\theta\delta_u=-\tilde g_\theta$, the score centered under the sampler. Since $\mathbb{E}_{\mu}\tilde g_\theta=0$, state-only terms drop and the return $R$ replaces $A^{\mu}$:
$\tilde g_\theta(s,a)=g_\theta(s,a)-\mathbb{E}_{v\sim\mu(\cdot\mid s)}\,g_\theta(s,v),\qquad g_\theta(s,a)=\nabla_\theta\log\pi_\theta(a\mid s)$
$\nabla_\theta\mathcal{L}=-\mathbb{E}_{\mu}\textstyle\sum_u\big(R-\beta\,\ell_{\theta,u}\big)\,\tilde g_\theta(s_u,a_u)\;\;\xrightarrow{\;\beta\to 0\;}\;\;-\mathbb{E}_{\mu}\textstyle\sum_u R\,\tilde g_\theta(s_u,a_u)$
4. The $\beta\to 0$ limit is score centering: the return times the sampler-centered score.

$\mu$: sampler (reference) policy · $c_f$, $c_u$, $c$: log-partition terms · Score centering: Marek & Ryabinin (2026).

Pillar II

Model Architectures

Structure the state — expressive and efficient.

The problem

Attention is memory-bound

  • 01The KV cache grows linearly with context — the wall for long-context and inference.
  • 02Compute is quadratic in sequence length.
  • 03Today's fixes (MQA / GQA) trade quality for memory.

What if the state itself were low-rank?

NeurIPS 2025 · Spotlight

Tensor Product Attention

Factorize queries, keys, and values as contextual tensor products — compressing the KV cache by up to ~10×.

$\;Q_t=\sum_{r=1}^{R}\,a^{Q}_{t,r}\otimes b^{Q}_{t,r}\,,\;$ likewise for $K_t,\,V_t$ — rank $R \ll d$.
  • →Longer context at fixed memory; unifies MHA / MQA / GQA as special cases.
  • →Compatible with any positional encoding (RoPE, ALiBi, additive RPE, GRAPE); the T6 backbone trains better at equal budget.
Tensor Product Attention: queries, keys and values are each built from low-rank factor pairs, with RoPE applied to one factor, then fed to scaled dot-product attention.

"Tensor Product Attention Is All You Need."

Pillar II · Tensor Product Attention · experiments

Lower loss with a much smaller KV cache

  • 01Lower validation loss. TPA and TPA-KVonly stay below MHA at nearly all training stages; MLA trains more slowly.
  • 02Better downstream. At 353M, TPA averages 51.41% zero-shot, against 50.11% for MHA, 50.44% for MQA and 50.13% for MLA.
  • 03Holds at scale. TPA-KVonly has the highest zero-shot average at 773M (53.52%) and at 1.5B (55.03%).
  • 04Up to ~10× smaller KV cache at equal quality.
Training and validation loss of large models on FineWeb-Edu 100B for MHA, MQA, GQA, MLA, TPA-KVonly and TPA: the two TPA curves end lowest and MLA highest.

NeurIPS 2025 Spotlight · 460 GitHub stars · Curves: large (773M) models on FineWeb-Edu 100B. Averages: zero-shot, lm-evaluation-harness.

Pillar II · GRAPE · ICLR 2026

Positions as group representations

Positions act via one-parameter subgroups, giving the General Relative Law: $G(t-s)=G(s)^{-1}G(t)$, so scores depend only on the offset $t-s$.
GRAPE-M · Multiplicative · SO(d) nω G(n) = exp(nωL) norm-preserving rotations,rank-2 skew generators L recovers RoPE exactly; extends to learned commuting subspaces and non-commuting mixtures GRAPE-A · Additive · GL(d+k) InA0I unipotent lift with A² = 0:position enters as anadditive logit bias recovers ALiBi and FoX exactly, with the exact relative law preserved

One frame, exact relative position laws, streaming cacheability · see also Thinking Machines' Inkling and Jane Street's blog

Pillar II · GRAPE · experiments

A persistent edge over RoPE, ALiBi and FoX

  • 01Controlled comparison. Llama-style models on FineWeb-Edu 100B; only the positional encoding changes.
  • 02Medium (353M). GRAPE-AP averages 53.25 on seven zero-shot tasks, against 51.73 for RoPE, 52.87 for ALiBi and 52.96 for FoX.
  • 03Large (770M). GRAPE-AP averages 56.91, against 55.76, 56.44 and 56.30.
  • 04Steadier training. At 770M the RoPE model shows training instability, while GRAPE improves steadily.
Training and validation loss of medium models on FineWeb-Edu 100B for RoPE, ALiBi, FoX and GRAPE-AP: RoPE stays highest through training and GRAPE-AP ends lowest in training loss.

GRAPE · arXiv:2512.07805 ↗ · Curves: medium (353M) models. Averages over ARC-E, ARC-C, HellaSwag, OBQA, PIQA, WinoGrande and SciQ, without KV-shift.

Pillar II · Recurrent Looped Transformer · latent reasoning

Latent reasoning: computation carried across tokens

A causal encoder supplies memory while the recurrent decoder passes its final state to the next token.

Feedback in the hidden state. The decoder’s final state enters the next token’s update; prompt and response use the same transition.

Depth grows with length. After $t$ tokens the recurrent path spans $tL_D$ decoder blocks, while each token still costs $L_E+L_D$.

100%parity at 256 bits, trained on at most 40 · Transformer 50.07%
97.30%S₅ swaps at 256 operations, 8× the training length · Transformer 0.85%
93.36%flat mod-5 at length 63 · Transformer 33.20%

RLT · arXiv:2610.07591 ↗ · Full RLT talk ↗ · Three-seed means; splits 5+3 and 7+1 (parity), 4+4 (swaps), 6+2 (mod-5), against an eight-layer Transformer.

Pillar II · RLT · reinforcement learning

Zero training–inference mismatch for stable RL

  • 01One transition. Prompt, response, user and tool tokens all take the same update; the prompt–response boundary does not reset state or caches.
  • 02Exact replay. After each parameter update, the trainer rebuilds encoder KV, recurrent states and decoder caches under the current parameters.
  • 03Fixed behavior probabilities. The sampler’s recorded log-probabilities stay as the ratio denominator.
  • 04Same execution settings. Replay uses the rollout’s chunk size, masks and positions.
Replaying a sampled history under current parameters: the sampler records tokens and behavior log-probabilities; the trainer rebuilds encoder KV, states and decoder caches, then forms the ratio of current to recorded probabilities.

RLT · arXiv:2610.07591 ↗ · Replay rebuilds current-policy state and caches; the recorded sampler probabilities remain fixed.

Pillar II · RLT · hardware efficiency

Hardware-efficient: parallel wherever tokens are known

  • 01Parallel encoder. Known tokens build the causal global KV memory together.
  • 02Chunked feedback (RLT-2). One feedback state per chunk of $B$ tokens cuts sequential decoder stages from $TL_D$ to $\lceil T/B\rceil L_D$, with the same parameters.
  • 03Batched across sequences. Independent sequences share one batched kernel for their next decoder update.
  • 04$B$ trades depth for parallelism. $B=1$ gives the longest recurrent path; a larger $B$ puts more known tokens in each parallel step.
Decoder batches for eight known tokens: RLT-0 runs one batch, RLT-1 runs eight token batches, and RLT-2 with B = 4 runs two chunk batches that share a feedback state.

RLT · arXiv:2610.07591 ↗ · The interval is a tradeoff: at $B=4$, 64-bit parity stays at 98.99% while S₅ swaps at 64 operations fall from 100% to 19.60%.

Pillar III

AutoResearch

Agents that automate research itself.

Pillar III · Agora · collective AutoResearch

Git records the research process

Agora mechanisms: Git storage, evidence propagation and UCB attention allocation.
  • 01Append-only contributions. Code, hypotheses, results and verifications form a dependency graph that workers can check out and reproduce.
  • 02Evidence guides attention. Reuse and verification propagate scores; analyze() exposes frontier results, contested claims and underexplored branches.

Agora · arXiv:2609.18094 ↗ · NeurIPS 2026 AutoMLR Workshop Oral · NVIDIA Technical Report
Workers choose experiments from shared state; there is no central planner assigning tasks.

Pillar III · Agora · sustained research run

Collective search leaves reproducible evidence

Agora weight-transfer study: development-evaluator bits per byte improves across the shared research graph.

141 donors, 32 architecture families. Transfer into a 119.6M-parameter attention–SSM hybrid using donor weights and forward passes.

3.3923 → 1.899 bpb. The development evaluator scores 200 FineWeb-Edu texts; pretraining and fine-tuning are prohibited.

13workers · nearly 12 days
1,703contributions
165independent reproductions · 0 failures

Agora · arXiv:2609.18094 ↗ · One weight-transfer study; bpb is the task’s development-evaluator score.

What's next

Toward autonomous superintelligence

  • 01AutoResearch: automate AI research itself, so improvements compound instead of being hand-built.
  • 02Agentic RL scaling: single-rollout learning with FlashREINFORCE.
  • 03Million-token memory: architectures for long context and continual learning.
  • 04Collective AutoResearch: use Agora to share experiments, verify results, and allocate the next research effort.

Principled methods, built to ship.

Thank you

Thank you.

With gratitude to my collaborators and advisors.

← Back
1 / 53

Slide navigatorClick a slide to jump · Esc to close

Keyboard & controls

→Space↓
Next build / slide
←↑
Previous
HomeEnd
First / last slide
F
Toggle fullscreen
O
Slide overview
S
Toggle fast mode (on by default) — off steps through builds
?H
This help
Esc
Close overlay

Press any key to close