h t t p : / / a c t - r . p s y . c m u . e d u / w o r d p r e s s / w p - c o n t e n t / u p l o a d s / 2 0 1 2 / 1 2 / 5 1 5 u i r - 2 0 0 4 - 0 7 - p i r o l l i . p d f

Ingest: The Use of Proximal Information Scent to Forage for Distal Content on the World Wide Web

Type: kb/sources/types/ingest-report.md

Classification

A chapter for an edited academic volume (Kirlik, ed., Working with Technology in Mind), issued as PARC tech report UIR-2004-07. It states a theory, derives a mechanism from a Bayesian rational analysis, and reports an empirical fit against coded think-aloud protocols, so it carries a paper's evidential obligations rather than an essay's. It is a synthesis of already-published work (Pirolli & Card 1999; Card et al. 2001; Morrison et al. 2001; Pirolli & Fu 2003) rather than a first report of new results.

Author: Peter Pirolli, PARC — originator of Information Foraging Theory with Stuart Card. That is the strongest credibility signal available on this topic and also the main bias to hold: the chapter is the program's author summarising his own program, and it reports no result that disconfirms it.

Summary

Pirolli frames Web navigation as a Brunswikian lens-model judgment problem: the forager never observes the distal content, only proximal cues (link summaries, search-result snippets, node labels, thumbnails), and must predict the one from the other in a probabilistically textured environment. He argues the cues have real ecological validity — Davison's elaborated anchor text correlates with the page it links to at r = .16 against roughly 0 for a random page, and page-to-page similarity falls sharply with link distance, so the Web has topical patches. He then derives the forager's prediction problem as Bayesian log-odds, maps it onto ACT-R spreading activation (Ai = Bi + Σj Wj Sji, with association strengths estimated from corpus co-occurrence and shown equal to PMI), and closes it with a Random Utility Model choice rule (multinomial logit) so that choice is stochastic rather than always-take-the-best. The resulting SNIF-ACT simulation is checked against 189 coded actions from four participants on two of six representative-design tasks: higher-scent links are the ones followed, and model-computed scent declines across a run of pages before the user abandons the site, with a high-scent start page predicting a longer run. Read it if you want the canonical mechanism account of judging an unseen target from its pointer, plus a stopping rule derived from the same quantity; skip the SNIF-ACT architecture sections if you only need the framing and the cue-validity measurements.

Quotes

  • Source extract (verbatim): scent refers to the cues used by information foragers to make judgments related
  • Source location: Introduction, PDF page 3.
  • Source extract (verbatim): to the selection of information sources to pursue and consume. These cues
  • Source location: Introduction, PDF page 3.
  • Source extract (verbatim): concise information about content that is not immediately available. The
  • Source location: Introduction, PDF page 3.
  • Source extract (verbatim): assessments of proximal information scent cues in order to make action choices
  • Source location: Introduction, PDF page 4.
  • Source extract (verbatim): that lead to distal information sources. This view is a variant of Brunswik’s Lens

  • Source location: Introduction, PDF page 4.

  • Source extract (verbatim): Human-information interaction systems will tend to maximize the

  • Source location: Adaptation framework and Equation 1, PDF page 5.
  • Source extract (verbatim): value of external knowledge gained relative to the cost of
  • Source location: Adaptation framework and Equation 1, PDF page 5.
  • Source extract (verbatim): (the anchor plus additional surrounding text, having a mean of 11.02 terms) to a
  • Source location: Topical Patches and Diminishing Returns, PDF page 10.
  • Source extract (verbatim): Linked r = .16 and Random r ≈ 0. Davison’s analysis of the correlation of
  • Source location: Topical Patches and Diminishing Returns, PDF page 10.
  • Source extract (verbatim): proximal cues to distal content confirms our intuition that the cues have
  • Source location: Topical Patches and Diminishing Returns, PDF page 10.
  • Source extract (verbatim): The stronger the associations (reflecting greater predictive strength)

  • Source location: Mapping the Bayesian Rational Analysis to Spreading Activation, PDF page 14.

  • Source extract (verbatim): Information scent refers to the cues used by information foragers to make judgments related to the selection of information sources to pursue and consume. These cues include items such as Web links or bibliographic citations that provide users with concise information about content that is not immediately available.

  • Source location: Introduction, PDF pages 3–4
  • Source extract (verbatim): Davison (2000) compared elaborated anchor text (the anchor plus additional surrounding text, having a mean of 11.02 terms) to a paired document that was either Linked (the page linked to the anchor) or Random (a random page). The normalized correlation (cosine) similarities were Linked r = .16 and Random r ≈ 0.

  • Source location: “Topical Patches and Diminishing Returns,” PDF page 10

  • Source extract (verbatim): Human-information interaction systems will tend to maximize the value of external knowledge gained relative to the cost of interaction.

  • Source location: Adaptation framework and Equation 1, PDF page 5
  • Source extract (verbatim): From proximal information scent cues, the user predicts the utility of a distal information source and makes choices based on the utilities of alternatives.
  • Source location: “The Random Utility Model,” PDF page 15
  • Source extract (verbatim): Because of the stochastic nature of the utilities UK|G it is not the case that one alternative will always be chosen over another.
  • Source location: “The Random Utility Model,” PDF page 16

Connections Found

The source is a primary account of the human proximal-cue/distal-content structure used in the KB's navigation cluster. Agents navigate by deciding what to read next now cites it as a narrower human analogue and separately states the local follow/skip abstraction, LLM transfer, and token/tool-call cost boundary. The chapter supports cue-based source selection, but it does not establish that follow/skip is the fundamental unit of navigation or compose cue quality and interaction cost into the target's pointer-level trade-off. Its strongest additional role is to ground and bound the human side of linking theory: the chapter supplies a value-over-interaction-cost tendency, an activation-plus-logit operationalization, and a protocol fit, while the note changes the priced resource to bounded LLM context. The source also sharpens link-following and search by giving a causal direction to its two-mode split, and it supplies a bounded human case of cue-based candidate discovery for the end-to-end knowledge-access synthesis. It does not establish that synthesis's LLM-agent transfer or whole-path conclusion. Two notes gain captured prior work in place of recalled prior work: information value is observer-relative, whose utility term is goal-conditioned by construction here, and soft-bound traditions, whose twelve-row table omits the tradition that fits its own (processor, bound, mechanism) schema most directly. The countervailing role is a limitation: the population is 14 human students on the 1998–2003 Web paying wall-clock time, so every edge into an LLM-agent claim needs the transfer argument that human-LLM differences are load-bearing demands, and this source cannot supply it.

Extractable Value

  1. A name and a measure for cue-to-content validity. The KB has vocabulary for a pointer "carrying context" but none for how well a cue predicts what it points at; ecological validity is that term, and Davison's anchor-text correlations make it a measurable property of a description rather than a stylistic preference. This is the concept a description-quality gate would need to state its own criterion. [quick-win]

  2. Pointer quality and stopping are one quantity, not two design problems. The same scent estimate both ranks the next move and sets the threshold for abandoning a patch. The KB reasons about pointer form and about when discovery stops in separate tags; the chapter's rate model (expected value of knowledge gained over cost of interaction) is the thing that would join them. [deep-dive]

  3. Cue quality causes mode switching. Weak-scent tasks pushed participants out of link-following and into search (about 3.25 versus 1.25 result sets per participant), and higher-rated links were the ones actually followed. Our two-mode split treats link-following and search as parallel regimes with different metadata needs; this says search is the fallback that fires when link cues fail, which makes description quality the lever on how often the expensive mode runs. [quick-win]

  4. A transferable measurement method. Web Behavior Graphs — a state-transition graph over a coded session, shaded by independently rated scent — plus the link/keyword/URL problem-space partition were produced from logged sessions with 91% and 93% inter-coder agreement. The same construction over logged agent sessions would let us measure our own navigation instead of reasoning about it analytically. [experiment]

  5. A thirteenth row for the soft-bound traditions table. Processor: information seeker. Bound: time and attention in a probabilistically textured environment. Adaptation mechanism: scent-guided patch selection plus a patch-leaving threshold. It is closer to that note's subject than several rows already present. [quick-win]

  6. A goal-conditioned utility that is observer-relative by construction. U_{J|G} is defined relative to the forager's goal and to association strengths learned from that forager's own history, so the same cue has different value for different observers by derivation rather than by assertion. That is a captured prior-work entry for a note whose survey is currently flagged as recalled. [just-a-reference]

Limitations (our opinion)

Our reading, not the author's.

The empirical baseline is a floor, not a comparison. Link choice is tested against random selection. Nothing in the chapter pits the spreading-activation-plus-RUM scent measure against a simpler rival, even though the obvious rival — cosine similarity over anchor text — is computed a few pages earlier for Davison's validity numbers, and an LSA-based alternative (Blackmon et al. 2002) is cited in the discussion. The fit therefore supports "some cue-similarity measure predicts navigation choice," not "this derivation is the right account of it."

Nothing was ablated, and the model's update space is narrow. The only thing estimated from data is the association strengths, taken once from Tipster plus AltaVista counts. Fixed outside that space: cues are bags of words, the goal is a fixed word set, utility is summed activation, the choice rule is Gumbel, the link/keyword/URL partition came from analyst hand-coding, and the scent ground truth behind the patch-leaving figures is human judges' ordinal ratings. There is no within-session learning at all — the chapter concedes that different foraging histories should yield different category structures and files that under future work. So the reported fit exercises choices inside a decomposition it cannot itself test, and none of it attributes the result to the Bayesian derivation over any other scoring function (learning inside a fixed decomposition inherits its mistakes).

The numbers are small where the claims are strongest. Fourteen participants were run, four protocols on two of six tasks were analysed, 189 actions were coded, and the site-leaving curve averages twelve site-leaving actions. The related figure of judge-rated scent before departure is explicitly ordinal, with the author warning against reading it quantitatively. The patch-leaving threshold is the chapter's most reusable claim and rests on its thinnest data.

Scope conditions are dated and human. Stanford students, the 1998–2003 Web, hint scaffolding at ten and fifteen minutes, and cost measured in wall-clock time. The part most likely to transfer to an agent is the structural one — judge the target from its pointer and never load it — because that is what makes navigation cheap for any bounded reader. The part least likely to transfer is the learned-strengths mechanism: our reader pays linear context per byte with no skim, and starts each session with no association strengths at all. Treat any edge from this source into an LLM-agent claim as owing that argument rather than inheriting it.

Self-report on a research program. Roughly every supporting result is the author's own, and the failure cases discussed (non-text cues, learning) are framed as open problems rather than as evidence against the framework.

Use this ingest as the bounded human baseline when adjudicating link-following and search. If the mode-switching use survives, retain its exact source passage or mark the route (snapshot required). That note's document-surrogate and search-result claims still need their separately assigned sources before its artifact disposition can be decided.