weekend paper reading: w46

17 Nov, 2025

I lied this was done on monday. very incomplete.

==================================

trying something out

1. first i read all the papers in the list, get their ideas, note it down
2. then i use a lite version of this framework to think more critically about each paper (summarize -> strengths -> weaknesses -> improvements/questions)

i've been thinking about how to be more critical reading papers and not be like this. this comment really stuck with me:

This reviewer doesn't understand at all that a research paper only needs to clearly present and justify its main contribution, but instead asks for a product development report where every microscopic detail must be thoroughly studied.

my current understanding is that a paper needs to 1) present its points and 2) justify it sufficiently. criticism should be focused on

• quality of idea and methodology
• whether the idea is sufficiently justified with the evidence provided
• how it compares to previous works (and whether it overlooks any)

pretraining large language models in nvfp4

https://arxiv.org/pdf/2509.25149

notes

smaller number = faster computation mixing precisions, numerically sensitive layer kept in high precisino improved quantization scheme using

• block quantization
  • values are scaled in blocks instead of individually
  • preserves consistency across transformations in forward and backward
  • scale is shared within the block
    • with 1D scaling you're doing max(row) / num_vals, so the scale is different for forward and backwards
    • with 2D you're using a whole block so it's the same even after transposition
  • less granular but just let the weights adapt
• random hadamard transforms
  • this means using random hadamard matrices to limit outliers
  • hadamard matrices are orthogonal matrices with values limited to ±1
  • H_d = (1/√2) H_2 ⊗ H_{d/2}
  • redistributes weights, but preserves energy
  • deterministic because of seed
  • orthogonal means they cancel out

takeaways

• so this paper shows that that it's possible to achieve near equal performance to fp8 with nvfp4
  • nvfp4 ~matches fp8 in perf on a 12B model with 10T tokens
  • matches fp8 entirely if you keep the last 20% of training in bf16/fp8
• mixed precision is good, keeping the final few layers in bf16 shows better perf
• combined
  • nvfp4
  • hadamard random transform
  • 2d block scaling
  • mixed precision

nvfp4 format

• block size 16 elements
• E4M3 format instead of UE8M0, greater range of expression

what is a block

1. fp32 mapped into representable range of a block
2. per block e4m3 scale moves value into block

Virtual Width Networks

https://arxiv.org/pdf/2511.11238

wider embedding -> send 2/3 of that into the transformer, remaining third is used as weighted rescons

notes

• decouples representational width from backbone width (?)
  • big embedding space
• reportedly scales well (2x for NTP and 3x for N2TP)
• more compression?
  • yeah the embedding is compressed before being passed into the transformer
• bigger embedding, replace residual connection with hyperconnections
  • bigger embedding - more representations
• hyperconnections?
  • residual connections with weights basically. skimmed https://arxiv.org/pdf/2409.19606 in like a minute but yeah
  • they route the connection matrices based on the inputs

contributions

• vwn
• generalized hyperconnections
• demonstrates that this works with MTP

MarsRL: Advancing Multi-Agent Reasoning System via Reinforcement Learning with Agentic Pipeline Parallelism

https://arxiv.org/pdf/2511.11373

pipeline for multi agent reasoning system

notes

• read gemini imo gold paper for source
• prev. solver-verifier-corrector pipeline:
  • each agent is given a single purpose
  • solver solves the problem
  • verifier detects error
  • corrector fixes accordingly
  • this technique didn't work with OSS models as well
• much RL is tool-integrated reasoning
  • issues with current RL
    • rewards are shared

marsRL

VC RL works great, but apparently only for Gemini 2.5 Pro.

• rewards are assigned to agents individually, credit roles are decoupled
  • solver and corrector gets rewarded based on their solutions and ground truth
  • verifier gets rewarded based on the correctness of its classification; if correct answer + incorrect verifier -> verifier gets penalized
    • ain't this very similar to a GAN
• pipeline parallel on an agent level?
  • i assume this is t1(agent1_{task1}) t2(agent1_{task2} agent2_{task1}) etc etc

perf wise this improves qwen3 a3b thinking 2507 by 5% at least across the board, which is quite good. i'd like to see bigger models benchmarked with this though.

thoughts

• how does one engineer a system/framework for multi-agent RL for this
  • https://github.com/TsinghuaC3I/MARTI
  • https://github.com/PRIME-RL/TTRL
• hmm this is tree search basically
  • not quite monte carlo?
  • naturally one would wanna restrict the search space
  • intelligent way of doing it?
  • would be good for flexible workloads
  • hmm
• can you use similar pipelines to improve codegen or whatever?
  1. normal interactions, use it
  2. takes the user-agent interactions, then another model looks at that to identify what the model should have done for a better trace
  3. finetune using trace

Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline

https://arxiv.org/pdf/2507.15855

half of this paper was prompt lol

most of it was covered in the prev. paper, but it demonstrated that this technique worked to get gold. nice.

Photon: Federated LLM Pre-Training

https://arxiv.org/pdf/2411.02908

cross silo FL for global scale training

• small batch size + high LR
• async trianing
• 20% higher throughput compared to centralized distributed training? what? what centralized? how centralized? how are the GPUs linked together?

basics

• federated learning benefits
  • data locality
  • distributed gpus

thoughts

• woah distributed global serverless cluster