Keyword Harnesses Fail Open: A Cheap Diagnostic Ladder for Tool-Use Claims in Small Language Models

Hugging Face Daily Papers Papers

Summary

The paper shows that keyword-matching benchmarks can credit small language models with tool use they never perform, documenting a false positive in matched-architecture Spanish security LMs and proposing a cheap diagnostic ladder plus a targeted SFT recipe that repairs the failing 1B model.

Keyword-matching benchmarks can credit small models for tool use they never perform. We document such a false positive in a matched-architecture pair of Spanish security language models and propose a ladder of strict, cheap diagnostics. A 661.6M parameter model (approx. 65% code/technical text; no dedicated SFT) and a 1,109M model (web-heavy multi-phase curriculum; 6B-token tool-SFT) share decoder, tokenizer, and special tokens, scoring almost identically on lenient tool-use metrics (B4: 0.660 vs. 0.650). Verbatim-reproduction checks on training examples separate them completely: the 600M emits valid tool calls with generalized arguments on 6/6 examples; the 1B does so on 0/6 across checkpoints. A first-token probe localizes the 1B's failure to a missing prior (prob. 10^{-4}--10^{-5} on <|tool_call|>), which was erased by its web-heavy training phase. A targeted SFT recipe (diverse corpus, 5x higher learning rate, 2,202 steps, ~3.3 GPU-hours) repairs the 1B using three orders of magnitude fewer tokens than the failed phase. On all 269 corpus rows, valid emission rises from 0.100 to 0.959 (600M: 0.926). On 238 unseen prompts, the repaired 1B passes 0.536 vs. the 600M's 0.428 (p = 0.004). Embedding-drift checks show the repair did not move the trigger token's tied embedding (97.7% of the bf16 table remains bit-identical), meaning changes live in the surrounding network. Both models over-trigger, rarely answering negative prompts without a call (0.09 for 600M, 0.17 for repaired 1B). Factorial analyses confirm all repair configurations install the format, though suppression benefits from a diverse corpus remain a hypothesis due to seed sensitivity. This cheap diagnostic ladder costs minutes of CPU time and should gate tool-use claims on small models.
Original Article
View Cached Full Text

Cached at: 10/02/26, 08:31 PM

Paper page - Keyword Harnesses Fail Open: A Cheap Diagnostic Ladder for Tool-Use Claims in Small Language Models

Source: https://huggingface.co/papers/2610.02142

Abstract

Keyword-matchingbenchmarkscancreditsmallmodelsfortoolusetheyneverperform.Wedocumentsuchafalsepositiveinamatched-architecturepairofSpanishsecuritylanguagemodelsandproposealadderofstrict,cheapdiagnostics.A661.6Mparametermodel(approx.65%code/technicaltext;nodedicatedSFT)anda1,109Mmodel(web-heavymulti-phasecurriculum;6B-tokentool-SFT)sharedecoder,tokenizer,andspecialtokens,scoringalmostidenticallyonlenienttool-usemetrics(B4:0.660vs.0.650).Verbatim-reproductionchecksontrainingexamplesseparatethemcompletely:the600Memitsvalidtoolcallswithgeneralizedargumentson6/6examples;the1Bdoessoon0/6acrosscheckpoints.Afirst-tokenprobelocalizesthe1B’sfailuretoamissingprior(prob.10^{-4}--10^{-5}on<|tool_call|>),whichwaserasedbyitsweb-heavytrainingphase.AtargetedSFTrecipe(diversecorpus,5xhigherlearningrate,2,202steps,~3.3GPU-hours)repairsthe1Busingthreeordersofmagnitudefewertokensthanthefailedphase.Onall269corpusrows,validemissionrisesfrom0.100to0.959(600M:0.926).On238unseenprompts,therepaired1Bpasses0.536vs.the600M’s0.428(p=0.004).Embedding-driftchecksshowtherepairdidnotmovethetriggertoken’stiedembedding(97.7%ofthebf16tableremainsbit-identical),meaningchangesliveinthesurroundingnetwork.Bothmodelsover-trigger,rarelyansweringnegativepromptswithoutacall(0.09for600M,0.17forrepaired1B).Factorialanalysesconfirmallrepairconfigurationsinstalltheformat,thoughsuppressionbenefitsfromadiversecorpusremainahypothesisduetoseedsensitivity.ThischeapdiagnosticladdercostsminutesofCPUtimeandshouldgatetool-useclaimsonsmallmodels.

View arXiv pageView PDFProject pageAdd to collection

Get this paper in your agent:

hf papers read 2610\.02142

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2610.02142 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2610.02142 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2610.02142 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Beyond APIs: Probing the Limits of MLLMs in Physical Tool Use

arXiv cs.CL

This paper introduces PhysTool-Bench, a benchmark for evaluating multimodal large language models' ability to recognize and plan the use of physical tools in real-world scenes. The authors find that even the best model identifies only 58.7% of tools and completes just 21.0% of queries end-to-end, revealing a two-level deficit in perception and functional commonsense.