Applied Papers

What could a research result change in the way a business makes decisions, tests a system, or runs a process? These notes start with that question, return to the original papers, and distinguish what was measured from what might transfer to practice.

The reading map draws on Ritvik Rastogi’s Papers Explained archive. Each entry pairs a contribution with a question worth testing in an organization. The linked explanations introduce the work; the articles below cite the primary research.

Published notes

applied research note · 17 Oct 2026

Deep Research Needs an Evidence Budget

Deep-research agents need explicit budgets for discovery, inspection, verification, and stopping before their reports can support business decisions.

Reading map

Evidence systems and research agents

These papers treat research as a sequence of search, reading, state management, verification, and stopping decisions. The business question is not whether an agent can produce a long report. It is whether the organization can inspect how the report acquired and tested its evidence.

Research contribution
Makes long research trajectories reproducible with a fixed corpus and explicit search, open, and find operations.
Business test
Can a due-diligence run be replayed after a decision is challenged?

Paper 639

OpenSeeker

Research contribution
Builds difficult multi-hop questions from web structure and compresses noisy search history.
Business test
Can training cases require real evidence traversal instead of keyword matching?

Paper 638

DRACO

Research contribution
Evaluates research reports across factual accuracy, coverage, presentation, and citation quality.
Business test
Does the acceptance test score the dimensions the decision owner actually needs?

Paper 637

DeepSearchQA

Research contribution
Tests exhaustive, multi-step search where entity resolution and stopping are part of the task.
Business test
Does the system stop because the answer set is complete or because its budget expired?

Paper 636

DR Tulu

Research contribution
Trains long-form research with rubrics that evolve as the policy discovers new evidence and failure modes.
Business test
Can evaluation keep discriminating after the system learns yesterday's checklist?
Research contribution
Combines agentic mid-training, post-training, automatic data synthesis, and a compressed working memory.
Business test
Which research state must remain verbatim, and which can be summarized safely?

Paper 634

PokeeResearch

Research contribution
Isolates reward, RL algorithm, data curation, and error recovery as separate design choices.
Business test
Is a weak result caused by the model, the training signal, the task mix, or brittle execution?
Research contribution
Shows a single agent can combine web search, page reading, and Python under a compact tool interface.
Business test
Is a multi-agent organization necessary, or is it hiding a simpler control problem?
Research contribution
Trains research behavior against the noisy live web rather than only a closed retrieval corpus.
Business test
Has the evaluation included blocked pages, stale snippets, and inconsistent sources?

Paper 631

ReSearch

Research contribution
Learns when to interleave reasoning and search without supervised reasoning traces.
Business test
Does the agent search to resolve uncertainty, or merely because a workflow tells it to?

Paper 630

Search-R1

Research contribution
Uses outcome-based RL to teach multiple real-time searches inside a reasoning trajectory.
Business test
Are search calls improving the answer enough to justify their latency and cost?
Research contribution
Separates supervised deployment behavior from RL for accuracy and efficient tool use.
Business test
Which requirements are hard guardrails, and which should be optimized from outcomes?
Research contribution
Finds that retrieval quality depends on the tool harness and presentation, not only the retrieval algorithm.
Business test
Has the team compared the simplest exact-search baseline inside the actual user workflow?

Evaluation, governance, and safety

This cluster is about proxies. Benchmarks, rubrics, verifiers, safety classifiers, and privacy filters can make quality measurable, but every proxy creates a new surface for overfitting or omission.

Paper 629

FRAMES

Research contribution
Combines retrieval, multi-hop reasoning, and factual response quality in end-to-end RAG questions.
Business test
Does a RAG test isolate retrieval failure from reasoning failure?

Paper 628

CRAG

Research contribution
Adds diverse question types and simulated web and knowledge-graph APIs to factual QA.
Business test
Does the test set include changing facts, unavailable facts, and false premises?

Paper 627

FanOutQA

Research contribution
Measures broad answer-set construction with human decompositions across many documents.
Business test
Can the system enumerate a market completely without duplicating the same entity?

Paper 626

MultiHop-RAG

Research contribution
Tests inference, comparison, temporal, and null questions that require several documents.
Business test
Can the system say that no supported answer exists after a multi-source search?
Research contribution
Measures how model answers collapse diverse human preferences into similar responses.
Business test
Is an ideation tool widening the option set or standardizing everyone on one safe answer?
Research contribution
Locates likely first errors in reasoning without step labels or final-answer verification.
Business test
Can process monitoring find where a workflow first went wrong before damage propagates?
Research contribution
Randomly hides rubric criteria during training so a policy cannot optimize one fixed checklist.
Business test
Does performance survive when auditors vary the inspection criteria?

Paper 596

Shieldstral

Research contribution
Frames multimodal moderation under configurable policies as binary question answering.
Business test
Can policy changes be versioned independently from the content classifier?
Research contribution
Uses the verifier's score distribution instead of a single discrete judgment.
Business test
Are borderline cases routed by calibrated uncertainty rather than a brittle pass or fail?
Research contribution
Distills rubric-conditioned behavior into a model that no longer needs a verifier on every call.
Business test
Which quality controls can be internalized, and which still require independent review?

Paper 579

POW3R

Research contribution
Reweights rubric criteria according to which ones still distinguish current outputs.
Business test
Is optimization effort spent on unresolved quality gaps rather than already-saturated metrics?
Research contribution
Shows that better rubric scores can coexist with worse broader quality.
Business test
What holdout review would reveal checklist compliance without actual improvement?

Paper 572

IFBench

Research contribution
Tests instruction following on new, programmatically verifiable constraints.
Business test
Does a workflow obey unfamiliar contractual formats or only templates seen in tuning?
Research contribution
Uses an efficient long-context token classifier to find and redact personal information.
Business test
Is privacy filtering measured for recall, over-redaction, context length, and local policy?
Research contribution
Builds data, a benchmark, and a judge for detecting critical errors in generated procedures.
Business test
Can an operations assistant be tested for unsafe omissions rather than fluent wording?
Research contribution
Compares probability and log-probability rewards across verifiable and open-ended reasoning.
Business test
Does the reward preserve meaningful differences, or collapse outcomes into a coarse binary signal?

Paper 555

IH Challenge

Research contribution
Trains and evaluates obedience to instruction priority across system, developer, user, and tool roles.
Business test
Can lower-trust content override policy or authorization in the deployed prompt stack?
Research contribution
Turns instance-specific, interpretable criteria into reinforcement-learning rewards.
Business test
Can domain experts inspect and amend the definition of a good result?

Training, curricula, and specialist agents

The common theme is allocation: which examples, traces, rewards, teachers, and specialist tasks deserve training compute. For a business, the practical unit is not the parameter count alone but the cost of acquiring a dependable capability.

Research contribution
Uses repeated natural-language proof generation, verification, and refinement without a formal prover.
Business test
When does an expensive generate-check-refine loop beat one fast answer?

Paper 624

GenCluster

Research contribution
Clusters diverse candidate programs before ranking and submission.
Business test
Can validation spend cover different solution families instead of near-duplicates?

Paper 623

Nemotron 3 CC

Research contribution
Joins problem curation, synthetic traces, SFT, RL, and test-time correction for competitive programming.
Business test
Which stage contributes measurable reliability in an agentic software pipeline?
Research contribution
Finds that partial or endpoint traces can retain much of the post-training value of full traces.
Business test
Are costly full process logs necessary, or would verified milestones carry the signal?
Research contribution
Reports stronger cross-task interference in sequential SFT than in sparse RL updates.
Business test
Does adding one capability erase another in the release regression suite?
Research contribution
Concentrates much of the measured RL gain in a small number of middle layers.
Business test
Can adaptation scope be narrowed without losing the target behavior?

Paper 610

TailSFT

Research contribution
Focuses SFT on sequences the model has not yet covered rather than repeatedly fitting easy cases.
Business test
Is the training queue reducing failure coverage or only average loss?

Paper 609

TinyGSM

Research contribution
Uses large synthetic math data plus a verifier to train small generation systems.
Business test
Can synthetic cases and automated checks substitute for scarce expert-written examples?

Paper 608

TinyStories

Research contribution
Constrains vocabulary and story structure so very small models expose learning behavior clearly.
Business test
Can a reduced-domain prototype test the learning pipeline before full-scale spend?

Paper 607

CLRS-Text

Research contribution
Expresses algorithmic execution traces as text across classical tasks.
Business test
Can process reasoning be evaluated against exact state transitions rather than prose quality?
Research contribution
Combines token-level teacher guidance with trajectory-level correctness.
Business test
Can dense coaching remain subordinate to the outcome that matters?
Research contribution
Extracts reusable procedural skills from agent trajectories and injects them as compact instructions.
Business test
Which repeated workflow knowledge belongs in a playbook instead of the model weights?

Paper 602

SOAR

Research contribution
Rewards a problem-generating teacher according to the student's improvement on hard cases.
Business test
Can test generation be optimized for learning value rather than novelty alone?

Paper 601

SCRL

Research contribution
Converts hard problems into verified subproblems ordered as a curriculum.
Business test
Can a complex business process be taught through auditable intermediate competencies?
Research contribution
Rewrites long correct reasoning traces into shorter supervision before training a student.
Business test
How much explanation cost can be removed without removing the decision signal?

Paper 597

Leanstral

Research contribution
Specializes general code agents for Lean proof work without a bespoke operating shell.
Business test
When should a shared agent interface host a high-assurance specialist?
Research contribution
Uses agreement among a model's own rollouts to create a corrective teaching signal.
Business test
Can internal consensus improve a system without hiding correlated mistakes?

Paper 594

TREK

Research contribution
Uses a short teacher-guided phase to expand the student's exploration before returning to RL.
Business test
Is a weak learner missing knowledge, or merely failing to visit useful solution modes?
Research contribution
Teaches from demonstrations through a privileged self-teacher while the student trains on its own rollouts.
Business test
Can new procedures be learned without a separate reward service or broad capability loss?
Research contribution
Gives one model privileged answer information in the teacher role and removes it in the student role.
Business test
What information may guide training but will not exist at production time?
Research contribution
Applies teacher feedback to sequences sampled from the student rather than only fixed teacher outputs.
Business test
Is the teacher correcting mistakes the deployed student actually makes?
Research contribution
Constructs a capability direction from contrasting weak models and transfers it to a stronger student.
Business test
Can cheap specialist signals improve a stronger general system without copying a weak model wholesale?
Research contribution
Uses controlled ablations to identify which instructions, teachers, filters, and task sources matter for agent data.
Business test
Has the data pipeline measured contribution by source rather than assuming more traces are better?
Research contribution
Combines curriculum SFT, multi-domain RL, and self-distillation in a compact reasoning model.
Business test
What capability floor is achievable under a strict deployment-size budget?
Research contribution
Separates diverse solution discovery in SFT from reinforcement of correct paths in RL.
Business test
Should exploration and reliability be optimized in different training stages?
Research contribution
Builds a large reasoning model with human-generated data and iterative empirical optimization.
Business test
Is model provenance compatible with the organization's licensing and audit requirements?

Paper 576

Aryabhata 2

Research contribution
Applies phased RL to a cleaned curriculum for competitive-exam reasoning.
Business test
Does a narrow expert model benefit from domain sequencing more than a mixed training pool?
Research contribution
Balances five reasoning domains with adaptive sampling and difficulty-aware length penalties.
Business test
Can the training scheduler prevent one high-volume domain from consuming the model?

Paper 559

Composer 2

Research contribution
Trains a coding agent in realistic software environments for long planning and multi-step execution.
Business test
Does coding evaluation include repository navigation, recovery, and end-to-end task completion?

Paper 557

Beyond Web

Research contribution
Rephrases targeted documents to make synthetic pretraining data denser and more diverse.
Business test
Is synthetic data adding useful coverage or merely restating the same source distribution?
Research contribution
Optimizes translated prompts for naturalness, cultural fit, and difficulty across languages.
Business test
Has a multilingual product localized the task itself, not only translated its words?
Research contribution
Combines multi-stage RL and on-policy distillation in an efficient mixture-of-experts model.
Business test
Which staged training costs buy reusable capability across math, code, and agents?

Paper 551

QED Nano

Research contribution
Trains a compact proof model with SFT, rubric rewards, and an iterative reasoning cache.
Business test
Can formal-quality output be decomposed into drafting, checking, and memory controls?

Architecture, context, and deployment efficiency

These papers change where computation and parameters live. Their business relevance is a cost ledger: training cost, serving cost, task-specific storage, latency, and operational flexibility are different quantities.

Paper 620

NoRA

Research contribution
Normalizes LoRA's down-projection to improve optimization while retaining exact mergeability.
Business test
Does a cheaper adaptation method remain stable across tasks and release merges?
Research contribution
Applies low-rank parameterization to feed-forward layers during pretraining, not only adaptation.
Business test
Is the organization reducing training compute or only task-specific parameter count?

Paper 615

YaRN

Research contribution
Extends RoPE context lengths with limited additional training.
Business test
Does the long-context claim survive the document lengths and positions in production?

Paper 614

NoPE

Research contribution
Challenges explicit positional encodings on length generalization tasks.
Business test
Is the architecture feature improving the target workload or persisting by convention?

Paper 613

ALiBi

Research contribution
Adds distance-dependent attention bias instead of positional embeddings.
Business test
Can longer inputs be supported without retraining the entire representation scheme?

Paper 612

RoPE

Research contribution
Represents absolute position through rotation while exposing relative distance in attention.
Business test
Which context-extension methods remain compatible with the deployed backbone?
Research contribution
Compares the mechanisms that give a parallel Transformer token order.
Business test
Is sequence order encoded in a way that matches extrapolation and latency requirements?

Paper 604

Kimi K3

Research contribution
Combines sparse activation, native multimodality, and very long context at large scale.
Business test
Which claimed capability requires the full model, and which can be served by a smaller route?

Paper 603

Kimi K2.5

Research contribution
Jointly trains text and vision for an open multimodal agentic model.
Business test
Does multimodality improve the actual workflow or add an expensive unused interface?
Research contribution
Converts a trained dense checkpoint into a sparse mixture-of-experts model.
Business test
Can sunk pretraining cost be reused without multiplying serving complexity?
Research contribution
Adds native audio, image, and video support to an efficient sparse backbone.
Business test
Which modalities need one shared model, and which require separate quality controls?

Paper 583

Mellum 2

Research contribution
Uses sparse experts for code editing, debugging, tool use, and conversational programming.
Business test
Does one coding model meet both interactive latency and agentic execution needs?

Paper 582

Mellum

Research contribution
Targets low-latency, multi-line IDE completion with a compact open model.
Business test
Is the product optimizing for completion speed rather than benchmark breadth?

Paper 573

EMO

Research contribution
Encourages domain-specialized experts that can be composed or deployed selectively.
Business test
Can unused capabilities be removed from a deployment without retraining the whole model?

Paper 571

Olmo 3

Research contribution
Releases models, data, training details, and checkpoints across general and reasoning variants.
Business test
Does openness reduce vendor risk enough to offset self-hosting responsibility?

Paper 568

Arcee Trinity

Research contribution
Mixes local and global attention with sparse experts and explicit load balancing.
Business test
Are throughput gains preserved under the organization's real request distribution?
Research contribution
Builds Indic-focused reasoning models with locally optimized tokenization.
Business test
Does regional language coverage require a different tokenizer and data supply chain?
Research contribution
Combines hybrid sequence layers, sparse experts, low precision, and multi-token prediction.
Business test
Which optimization lowers total serving cost after hardware and software constraints are included?

Data and representation

This group concerns what the system stores, compares, and retrieves. The practical concern is whether the representation preserves the distinctions needed by a search, classification, or multimodal workflow.

Research contribution
Maps text, image, video, audio, and PDFs into one multilingual space with variable dimensions.
Business test
Does one index improve cross-modal discovery without erasing modality-specific evidence?
Research contribution
Adds frozen modality encoders around a text embedding space and trains small connectors.
Business test
Can new document types be added without rebuilding the entire retrieval stack?
Research contribution
Uses compression as a proxy for data complexity when predicting scaling behavior.
Business test
Can a cheap complexity measure improve data and compute planning before a training run?
Research contribution
Combines normalized compression distance with nearest-neighbor classification.
Business test
Has the team tested a transparent non-neural baseline before adopting a larger classifier?

Paper 561

FinePhrase

Research contribution
Treats synthetic pretraining data as a recipe optimized through many controlled experiments.
Business test
Are synthetic-data costs tied to ablations that show which transformations help?
Research contribution
Combines distillation and contrastive objectives for compact, multilingual, quantization-tolerant embeddings.
Business test
Do retrieval tests include truncation, language mix, and the quantized production index?

How to read these notes

Every applied note identifies the system and question before presenting a result. It cites the original paper, states what the experiment measured, names a plausible business use, and explains where the result may fail to transfer. A benchmark result alone does not establish production reliability or business value.