NLP & Language
Research on NLP & Language: explanations, evidence, limitations and connections between papers.
57 results
Models solve fewer problems after a structural change makes them NP-complete
A paired coding benchmark finds models solve NP-complete problems less reliably than polynomial-time counterparts, even with more time.
DynaWeave turns static tool tasks into changing scenarios
DynaWeave turns static tool tasks into changing scenarios; trained Qwen3.5 models improved across three dynamic tests.
Models that solve best are not always the best task authors
Environment Duels finds that strong language model task solvers are not necessarily the strongest authors of useful challenges.
Why preference ranking can improve as text generation collapses
ECHO traces how shared hidden-state shifts can preserve response rankings while disrupting text generation in tested language models.
A clause-based test exposes conflicts with professional ethics rules
TenetBench tests answers against named professional rules; most evaluated model rows often conflicted with the rule on improper requests.
A document parser preserves visual cues agents need to find requirements
A formatting-aware parser helps an answering model recover author-marked requirements in a real-policy-document task.
ProVe speeds token generation by checking proposed blocks
ProVe checks multi-token proposals, matching greedy output while increasing measured generation speed in six Qwen instruction tests.
A coordinated system makes low-bit language models generate in blocks
BitMTP coordinates low-bit block drafting and runtime execution, raising measured generation speed while keeping tested mean scores close to baselines.
Agents can drop old cache states without rewriting their history
Cache engineering lets language-model agents drop old cached states; tests report lower cost and improved accuracy in a bounded setup.
A repaired teacher helps agents finish more tasks
A teacher-score repair method improved task completion across the tested text-agent models and environments.
Monsoon builds speech data that improves conversational recognition
Monsoon collects verified spontaneous speech; its Whisper models reduce errors on independently collected conversational tests.
TokenGate cuts long-prompt processing while preserving accuracy
TokenGate selects prompt tokens for later processing, reducing prefill time and memory without accuracy loss across five tested tasks.
A system helps visual agents gather and reuse evidence
OpenVLHarness helps visual agents gather and reuse tool evidence; Qwen3-VL-8B scored higher on average across 22 visual task datasets.
Measure the consequence before trusting a model signal
A byte model can be confident that the next symbol is a digit while missing the calculation that determines its value. A structured judgment can fit its required format while changing a program’s action. Several accurate observers can still fail together. These studies suggest a practical sequence: identify the decision, then test the output or behavior on which it depends. The relevant endpoint differs across computed answers, program decisions, and the benefit of acting with shared errors. Each finding belongs to its own tested setting.
Match cache changes to the cost your workload pays
A smaller cache can save stored memory, while selective reads can reduce the information fetched for each new query. Removing visual entries during answer generation changes a third cost. A quality score is useful for choosing among these approaches only if its test exposes the phase being changed. The studies point to a practical first step: identify the constrained resource, then test the corresponding method on the intended workload. Their classification, language-modeling, and visual-answer results describe different outcomes, so they do not yield a shared ranking.
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.
Match instruction edits to the claim they test
A robot may complete a tabletop task after its instruction is reworded. A guard reviewing an agent’s proposed tool call may also change its decision when a request or benchmark marker is edited. The same act of changing text serves different purposes in these cases. To assess a helpful rewrite, compare task success with a stated baseline. To assess sensitivity, hold the case fixed and measure the decision change. The reported effects answer those separate questions within their tested models and settings.
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.