Founding AI Research Engineer

Talantir; EUA; Teletrabalho; USD 90 000 - 200 000; Tempo integral

Resumo da oferta

Deep Reasoning Labs is building a deep reasoning layer for LLMs focused on long-horizon coding. This role owns core R&D and implementation of test-time scaling for coding, with emphasis on search/branching policies, verifiers, and SFT/RL training signals.

You will ship systems that run continuously, generate data, and measurably improve complex codebase tasks. Remote-first with US/Canada overlap; compensation includes meaningful equity.

Qualificações

• Strong implementation ability in Python + PyTorch, with good engineering hygiene.
• Demonstrated experience building inference-time search/reranking/verifiers for LLMs or sequence models.
• Comfort with data pipelines, eval harnesses, and running systematic experiments.
• Clear thinking about failure modes, measurement, and falsifiable hypotheses.

Responsabilidades

• Own core R&D and implementation of test-time scaling for coding: search/branching policies, verifiers/PRMs, SFT fine-tuning, RLVR-style training signals.
• Ship systems that run continuously, generate data, and measurably improve complex codebase tasks.
• Develop program analysis + transformations signals (AST/IR, dataflow) to guide reasoning loops.
• Evaluate and design experiments to quantify gains and establish baselines.

Conhecimentos

Python

PyTorch

Inference-time search

Data pipelines

Experimentation

Descrição da oferta de emprego

Company

Deep Reasoning Labs is building a deep reasoning layer for LLMs focused on long-horizon coding: 10–30 minute solve loops with branch exploration, parallel execution, and search strategies grounded in real tool/runtime feedback and deep code analysis signals. The goal is a system that reliably improves on complex codebase tasks with more test-time compute.

Role

You will own core R&D and implementation of test-time scaling for coding: search/branching policies, verifiers/PRMs, SFT fine-tuning, RLVR-style training signals, and rigorous evaluation. A major emphasis is integrating program analysis and code transformation signals (static + dynamic analysis) into the reasoning loop. This is a hands-on role: you will ship systems that run continuously, generate data, and measurably improve.

What you’ll work on

Test-time scaling + search for code generation and debugging:

• Best-first / beam variants, MCTS-like approaches, branch pruning, reranking, caching


• Tool- and execution-grounded loops (compile/test/run/linters/typecheck)


• Branch management policies: when to explore vs exploit, how to allocate compute, how to stop early

PRM / verifier systems:

• Step-level scoring and trajectory evaluation


• Training data generation, label quality control, calibration, and failure analysis


• Building “decision policies” that use verifier signals to steer branching/search or prune early.

Post-training strategy:

• Parameter-efficient fine-tuning (e.g., QLoRA) with SFT and related methods


• RL-style post-training driven by verifiable rewards (e.g., tests/compilation/static checks; methods such as GRPO-style updates when appropriate.)


• Use verifiable signals (tests/compilation/static checks) and preference signals where useful

Program analysis + transformations:

• Using signals such as AST/IR structure, type system feedback, build graphs, lint/static checks, and deeper dataflow/CPG-style analysis


• Applying automated code transformations and repair actions that are validated by compilation/tests

Evaluation + ablations:

• Build suites for long-horizon coding tasks and measure compute vs success curves


• Design experiments to isolate what truly drives gains

What success looks like (first ~90 days)

• End-to-end “deepthink for code” loop running on a real task suite with branching + execution feedback


• Clear win vs a strong baseline using controlled ablations


• A verifier/PRM stack that improves selection and branch allocation under compute constraints

Requirements (must-have)

• Strong implementation ability in Python + PyTorch, with good engineering hygiene


• Demonstrated experience building inference-time search/reranking/verifiers for LLMs or sequence models


• Comfort with data pipelines, eval harnesses, and running systematic experiments


• Clear thinking about failure modes, measurement, and falsifiable hypotheses

Nice-to-have

• Experience with any of: PRMs, reward models, preference learning, RLHF/RLAIF, RL for sequence generation


• Strong background in compilers/tooling, program synthesis, automated debugging, or code agents


• Familiarity with code analysis signals (AST, type systems, static checks, CPG/dataflow concepts)

Tech stack

Python, PyTorch, vLLM/SGLang or equivalent serving, distributed training/inference tooling (will be managed by the platform engineer), internal eval infrastructure.

Location / work model

Remote-first (US/Canada). Strong preference for overlap with Pacific Time. Periodic in-person sprints in SF are a plus.

Compensation

Market-competitive base (location-based) + meaningful founding-level equity.

Pay: $90,000.00 - $200,000.00 per year

Work Location: Remote