NLP & Language
Research on NLP & Language: explanations, evidence, limitations and connections between papers.
42 results
Non-English requests more often enter retrieval
A multilingual retrieval-agent evaluation finds higher tool admission for non-English harmful requests in a controlled URL setting.
Post-training strengthens an existing circuit for following cues
A mechanistic study finds that later checkpoints strengthen cue effects through an existing attention circuit in two training pipelines.
A token-budget method predicts when debate groups help
A token-budget method predicts debate group size from single-agent tests, reducing selection regret in tested models and tasks.
Language model search can spend less on predictable steps
Uncertainty-Aware Allocation reduces language model search effort when candidate steps look alike, with accuracy trade-offs in tested settings.
Small-model confidence helps routing on some multilingual tasks
A multilingual evaluation finds confidence routing helps reading and topic tasks, but falters on reasoning.
Choose a past signal that fits the task and its resource limit
A document can be compressed before anyone asks about it; a robot can learn a new task while still needing an old one. Reusing the past helps only if the retained signal supports the job ahead. Across four studies, the practical choice is where to place that signal and what it costs to keep or recreate. Their results favor task-relevant signals in their tested settings, while the studies measure different outcomes and resources.
Instruction changes can steer behavior or expose wording sensitivity
A changed instruction can help a system complete a task, or it can reveal that a decision depends on wording. A robot’s request, a stalled math solution, and a guard reviewing a tool call make those purposes easy to confuse. The studies offer a bounded way to separate them: test steering by the task outcome it seeks to improve, and test sensitivity by changing one input feature while keeping the case fixed. The result belongs to the model, task, and comparison that produced it.
Intern-Decision scores structured choices in one pass
Intern-Decision scores structured choices directly; its 4B model has a higher mean accuracy than Jev across seven evaluated suites.
Multi-SFT builds permissive instruction data across five languages
Models trained on Multi-SFT had higher multilingual test averages than official instruction-tuned models in six of nine comparisons.
Adding context to safety examples reduces cued misalignment
ContextLift reduced cue-triggered misalignment in three advice domains on one tested language model.
A method steers answers as it revises question context
EMO combines question-specific context revision and token steering, improving rule-scored accuracy over zero-shot generation across its tested grid.
Long-context midtraining raises selected coding scores
An open midtraining recipe improved selected post-SFT coding test scores on two Nemotron models, with results varying by test.
Strong models can review agent steps instead of writing each
SpecAgent has a larger model review grounded tool drafts, matching Strong-only success on tested Airline tasks while using less estimated memory traffic.
Math overlap rules can remove many distinct problems
An evaluation of math-data cleaning finds case-by-case adjudication better balances reliability and coverage on reviewed candidate pairs.
Read the evaluation protocol before trusting a benchmark score
Imagine seeing different accuracy scores for the same video model on the same benchmark. The model weights may be unchanged, yet the evaluator can choose different frames, process them differently, or score the response by a different rule. Across video questions, Taiwan knowledge questions, and dependency repairs, these studies show why a result needs its protocol beside it. A compact test can make a comparison less costly in its tested setting, while its task coverage still limits the claim. For a repair, success also depends on which checks the verifier runs.
Check what stays useful before reusing computation
A recommendation feature can outlive its usefulness, a matrix update can arrive with a new set of changed inputs, and a robot can see a scene that has just moved. Keeping prior work may save computation, but its value depends on what was kept and what changed. Across these studies, the useful first step is to identify the reusable work and its conditions, then measure the outcome that matters for that system. Removing a feature may raise serving throughput; retaining a matrix schedule may lose to a simpler update; refreshing fewer robot features may shorten an action request. Those are distinct findings in distinct settings.
Dedicated model leads on Indian language speech tests
A dedicated model leads on average in shared Indian-language speech tests, with exceptions on long read speech.
A speech model confines feature matching to protect English
A multilingual speech method confines feature matching to adapters, lowering word error rates in reported English and multilingual test groups.
A tokenizer reuses chat history to cut first-token delays
Delta Tokenizer reuses session text to cut first-token delay in tested long-context SGLang chat requests.
HybridAgent combines latent reasoning with tool use in text
HybridAgent combines latent reasoning with tool interaction; staged math tests show gains, while a proof bounds cross-turn memory claims.
JetSpec++ drafts ahead while the target checks current tokens
Across seven tests, JetSpec++ averaged 8.94 committed tokens per round versus DSpark’s 6.89 on Qwen3-8B at temperature zero and budget 16.
Sparse attention stays close to full attention at long context
JSSA selects context tokens directly, staying near full-cache language-modeling scores while speeding decoding in tested setups.
A Brazilian Portuguese model adds a tested safety step
A Brazilian Portuguese model adds safety training and refuses most harmful prompts in a 218-item evaluation.
Only two agent-training methods beat GRPO on average
A comparison of ten text-agent training methods finds average gains are limited and do not consistently outweigh training cost.